enable stable PD plugin in premake4.lua PyBullet

add a normalize in the setRotation axis/angle to be sure.
Add more code from the DeepMimic project, training doesn't work yet.

examples/pybullet/gym/pybullet_envs/deep_mimic/env/action_space.py

@@ -0,0 +1,6 @@
+from enum import Enum
+
+class ActionSpace(Enum):
+    Null = 0
+    Continuous = 1
+    Discrete = 2

examples/pybullet/gym/pybullet_envs/deep_mimic/env/env.py

@@ -0,0 +1,13 @@
+from abc import ABC, abstractmethod
+import numpy as np
+from enum import Enum
+
+class Env(ABC):
+    class Terminate(Enum):
+        Null = 0
+        Fail = 1
+        Succ = 2
+        
+    def __init__(self, args, enable_draw):
+        self.enable_draw = enable_draw
+        return

examples/pybullet/gym/pybullet_envs/deep_mimic/env/pybullet_deep_mimic_env.py

@@ -0,0 +1,251 @@
+import numpy as np
+import math
+from pybullet_envs.deep_mimic.env.env import Env
+from pybullet_envs.deep_mimic.env.action_space import ActionSpace
+from pybullet_envs.minitaur.envs import bullet_client
+import time
+import motion_capture_data
+from pybullet_envs.deep_mimic.env import humanoid_stable_pd
+import pybullet_data
+import pybullet as p1
+import random
+      
+      
+class PyBulletDeepMimicEnv(Env):
+    def __init__(self, args=None, enable_draw=False, pybullet_client=None):
+      super().__init__(args, enable_draw)
+      self._num_agents = 1
+      self._pybullet_client = pybullet_client
+      self._isInitialized = False
+      self.reset()
+    
+    def reset(self):
+      self.t = 0
+      if not self._isInitialized:
+        self._pybullet_client =  bullet_client.BulletClient(connection_mode=p1.GUI)
+        #self._pybullet_client  =  bullet_client.BulletClient()
+        
+        self._pybullet_client.setAdditionalSearchPath(pybullet_data.getDataPath())
+        z2y = self._pybullet_client.getQuaternionFromEuler([-math.pi*0.5,0,0])
+        self._planeId = self._pybullet_client.loadURDF("plane.urdf",[0,0,0],z2y)
+        #print("planeId=",self._planeId)
+        self._pybullet_client.configureDebugVisualizer(self._pybullet_client.COV_ENABLE_Y_AXIS_UP,1)
+        self._pybullet_client.setGravity(0,-9.8,0)
+        
+        self._pybullet_client.setPhysicsEngineParameter(numSolverIterations=10)
+        self._pybullet_client.changeDynamics(self._planeId, linkIndex=-1, lateralFriction=0.9)
+        
+        self._mocapData = motion_capture_data.MotionCaptureData()
+        motionPath = pybullet_data.getDataPath()+"/motions/humanoid3d_walk.txt"
+        #motionPath = pybullet_data.getDataPath()+"/motions/humanoid3d_backflip.txt"
+        self._mocapData.Load(motionPath)
+        timeStep = 1./600
+        useFixedBase=False
+        self._humanoid = humanoid_stable_pd.HumanoidStablePD(self._pybullet_client, self._mocapData, timeStep, useFixedBase)
+        self._isInitialized = True
+        
+        self._pybullet_client.setTimeStep(timeStep)
+        self._pybullet_client.setPhysicsEngineParameter(numSubSteps=2)
+        
+        
+        selfCheck = False
+        if (selfCheck):
+          curTime = 0
+          while self._pybullet_client.isConnected():
+            self._humanoid.setSimTime(curTime)
+            state = self._humanoid.getState()
+            #print("state=",state)
+            pose = self._humanoid.computePose(self._humanoid._frameFraction)
+            for i in range (10):
+              curTime+=timeStep
+              #taus = self._humanoid.computePDForces(pose)
+              #self._humanoid.applyPDForces(taus)
+              #self._pybullet_client.stepSimulation()
+            time.sleep(timeStep)
+
+      startTime = random.randint(0,self._humanoid._mocap_data.NumFrames()-2)
+      self._humanoid.setSimTime(startTime)
+      self._humanoid.resetPose()
+      #this clears the contact points. Todo: add API to explicitly clear all contact points?
+      self._pybullet_client.stepSimulation()
+      
+       	
+    def get_num_agents(self):
+    	return self._num_agents
+    	
+    def get_action_space(self, agent_id):
+    	return ActionSpace(ActionSpace.Continuous)
+    	
+    def get_reward_min(self, agent_id):
+    	return 0
+    
+    def get_reward_max(self, agent_id):
+    	return 1
+    	
+    def get_reward_fail(self, agent_id):
+    	return self.get_reward_min(agent_id)
+    
+    def get_reward_succ(self, agent_id):
+    	return self.get_reward_max(agent_id)
+    	
+    #scene_name == "imitate" -> cDrawSceneImitate
+    def get_state_size(self, agent_id):
+        #cCtController::GetStateSize()
+        #int state_size = cDeepMimicCharController::GetStateSize();
+        #                     state_size += GetStatePoseSize();#106
+	      #                     state_size += GetStateVelSize(); #(3+3)*numBodyParts=90
+        #state_size += GetStatePhaseSize();#1
+        #197
+        return 197
+
+    def build_state_norm_groups(self, agent_id):
+        #if (mEnablePhaseInput)
+	      #{
+		    #int phase_group = gNormGroupNone;
+		    #int phase_offset = GetStatePhaseOffset();
+		    #int phase_size = GetStatePhaseSize();
+		    #out_groups.segment(phase_offset, phase_size) = phase_group * Eigen::VectorXi::Ones(phase_size);
+        groups = [0]*self.get_state_size(agent_id)
+        groups[0] = -1
+        return groups	
+        
+    def build_state_offset(self, agent_id):
+        out_offset = [0]*self.get_state_size(agent_id)
+        phase_offset = -0.5
+        out_offset[0] = phase_offset
+        return np.array(out_offset)
+    
+    def build_state_scale(self, agent_id):
+        out_scale = [1]*self.get_state_size(agent_id)
+        phase_scale = 2
+        out_scale[0] = phase_scale
+        return np.array(out_scale)
+
+    def get_goal_size(self, agent_id):
+        return 0
+
+    def get_action_size(self, agent_id):
+     	  ctrl_size = 43 #numDof
+     	  root_size = 7
+     	  return ctrl_size - root_size
+     	  
+    def build_goal_norm_groups(self, agent_id):
+        return np.array([])
+        
+    def build_goal_offset(self, agent_id):
+        return np.array([])
+    
+    def build_goal_scale(self, agent_id):
+        return np.array([])
+        
+    def build_action_offset(self, agent_id):
+    	  out_offset = [0] * self.get_action_size(agent_id)
+    	  out_offset = [0.0000000000,0.0000000000,0.0000000000,-0.200000000,0.0000000000,0.0000000000,0.0000000000,
+    	    -0.200000000,0.0000000000,0.0000000000,	0.00000000,	-0.2000000,	1.57000000,	0.00000000,	0.00000000,
+    	    0.00000000,	-0.2000000,	0.00000000,	0.00000000,	0.00000000,	-0.2000000,	-1.5700000,	0.00000000,	0.00000000,
+    	    0.00000000,	-0.2000000,	1.57000000,	0.00000000,	0.00000000,	0.00000000,	-0.2000000,	0.00000000,	0.00000000,
+    	    0.00000000,	-0.2000000,	-1.5700000]
+    	  #see cCtCtrlUtil::BuildOffsetScalePDPrismatic and
+    	  #see cCtCtrlUtil::BuildOffsetScalePDSpherical
+    	  return np.array(out_offset)
+    	  
+    def build_action_scale(self, agent_id):
+    	  out_scale = [1] * self.get_action_size(agent_id)
+    	  #see cCtCtrlUtil::BuildOffsetScalePDPrismatic and
+    	  #see cCtCtrlUtil::BuildOffsetScalePDSpherical
+    	  out_scale=[ 0.20833333333333,1.00000000000000,1.00000000000000,1.00000000000000,0.25000000000000,
+    	    1.00000000000000,1.00000000000000,1.00000000000000,0.12077294685990,1.00000000000000,
+    	    1.000000000000,	1.000000000000,	0.159235668789,	0.159235668789,	1.000000000000,
+    	    1.000000000000,	1.000000000000,	0.079617834394,	1.000000000000,	1.000000000000,
+    	    1.000000000000,	0.159235668789,	0.120772946859,	1.000000000000,	1.000000000000,
+    	    1.000000000000,	0.159235668789,	0.159235668789,	1.000000000000,	1.000000000000,
+    	    1.000000000000,	0.107758620689,	1.000000000000,	1.000000000000,	1.000000000000,
+    	    0.159235668789]
+    	  return np.array(out_scale)
+    
+    def build_action_bound_min(self, agent_id):
+    	  #see cCtCtrlUtil::BuildBoundsPDSpherical
+    	  out_scale = [-1] * self.get_action_size(agent_id)
+    	  out_scale = [-4.79999999999,-1.00000000000,-1.00000000000,-1.00000000000,-4.00000000000,
+    	    -1.00000000000,-1.00000000000,-1.00000000000,-7.77999999999,-1.00000000000,	-1.000000000,
+	    	  -1.000000000,	-7.850000000,	-6.280000000,	-1.000000000,	-1.000000000,	-1.000000000,
+	    	  -12.56000000,	-1.000000000,	-1.000000000,	-1.000000000,	-4.710000000,
+	    	  -7.779999999,	-1.000000000,	-1.000000000,	-1.000000000,	-7.850000000,
+	    	  -6.280000000,	-1.000000000,	-1.000000000,	-1.000000000,	-8.460000000,
+	    	  -1.000000000,	-1.000000000,	-1.000000000,	-4.710000000]
+
+    	  return out_scale
+    
+    def build_action_bound_max(self, agent_id):
+    	  out_scale = [1] * self.get_action_size(agent_id)
+    	  out_scale=[
+    	    4.799999999,1.000000000,1.000000000,1.000000000,4.000000000,1.000000000,
+    	    1.000000000,1.000000000,8.779999999,1.000000000,	1.0000000,	1.0000000,
+    	  	4.7100000,	6.2800000,	1.0000000,	1.0000000,	1.0000000,
+	    	  12.560000,	1.0000000,	1.0000000,	1.0000000,	7.8500000,
+	    	  8.7799999,	1.0000000,	1.0000000,	1.0000000,	4.7100000,
+	    	  6.2800000,	1.0000000,	1.0000000,	1.0000000,	10.100000,
+	    	  1.0000000,	1.0000000,	1.0000000,	7.8500000]
+    	  return out_scale
+    	  
+    def set_mode(self, mode):
+    	  self._mode = mode
+    	  
+    def need_new_action(self, agent_id):
+        return True
+        
+    def record_state(self, agent_id):
+        state = self._humanoid.getState()
+        state[1]=state[1]+0.008
+        #print("record_state=",state)
+        return state
+        
+        
+    def record_goal(self, agent_id):
+        return np.array([])
+        
+    def calc_reward(self, agent_id):
+        kinPose = self._humanoid.computePose(self._humanoid._frameFraction)
+        reward = self._humanoid.getReward(kinPose)
+        return reward
+        
+    def set_action(self, agent_id, action):
+        self.desiredPose = self._humanoid.convertActionToPose(action)
+        #print("set_action: desiredPose=", self.desiredPose)
+        
+    def log_val(self, agent_id, val):
+        pass
+        
+    def update(self, timeStep):
+        for i in range(10):
+          self.t += timeStep
+          self._humanoid.setSimTime(self.t)
+          
+          if self.desiredPose:
+            kinPose = self._humanoid.computePose(self._humanoid._frameFraction)
+            self._humanoid.initializePose(self._humanoid._poseInterpolator, self._humanoid._kin_model, initBase=False)
+            pos,orn=self._pybullet_client.getBasePositionAndOrientation(self._humanoid._sim_model)
+            self._pybullet_client.resetBasePositionAndOrientation(self._humanoid._kin_model, [pos[0],pos[1]+2,pos[2]],orn)
+            #print("desiredPositions=",self.desiredPose)
+            taus = self._humanoid.computePDForces(self.desiredPose)
+            self._humanoid.applyPDForces(taus)
+            self._pybullet_client.stepSimulation()
+            
+
+    def set_sample_count(self, count):
+        return
+        
+    def check_terminate(self, agent_id):
+      return Env.Terminate(self.is_episode_end())
+      
+    def is_episode_end(self):
+      isEnded = self._humanoid.terminates()
+      #also check maximum time, 20 seconds (todo get from file)
+      print("self.t=",self.t)
+      if (self.t>3):
+        isEnded = True
+      return isEnded
+      
+    def check_valid_episode(self):
+      #could check if limbs exceed velocity threshold
+      return true

examples/pybullet/gym/pybullet_envs/deep_mimic/learning/agent_builder.py

@@ -0,0 +1,20 @@
+import json
+import numpy as np
+from learning.ppo_agent import PPOAgent
+
+AGENT_TYPE_KEY = "AgentType"
+
+def build_agent(world, id, file):
+    agent = None
+    with open(file) as data_file:    
+        json_data = json.load(data_file)
+        
+        assert AGENT_TYPE_KEY in json_data
+        agent_type = json_data[AGENT_TYPE_KEY]
+        
+        if (agent_type == PPOAgent.NAME):
+            agent = PPOAgent(world, id, json_data)
+        else:
+            assert False, 'Unsupported agent type: ' + agent_type
+
+    return agent

examples/pybullet/gym/pybullet_envs/deep_mimic/learning/exp_params.py

@@ -0,0 +1,54 @@
+import json
+import numpy as np
+import pybullet_utils.math_util as MathUtil
+
+class ExpParams(object):
+    RATE_KEY = 'Rate'
+    INIT_ACTION_RATE_KEY = 'InitActionRate'
+    NOISE_KEY = 'Noise'
+    NOISE_INTERNAL_KEY = 'NoiseInternal'
+    TEMP_KEY = 'Temp'
+
+    def __init__(self):
+        self.rate = 0.2
+        self.init_action_rate = 0
+        self.noise = 0.1
+        self.noise_internal = 0
+        self.temp = 0.1
+        return
+
+    def __str__(self):
+        str = ''
+        str += '{}: {:.2f}\n'.format(self.RATE_KEY, self.rate)
+        str += '{}: {:.2f}\n'.format(self.INIT_ACTION_RATE_KEY, self.init_action_rate)
+        str += '{}: {:.2f}\n'.format(self.NOISE_KEY, self.noise)
+        str += '{}: {:.2f}\n'.format(self.NOISE_INTERNAL_KEY, self.noise_internal)
+        str += '{}: {:.2f}\n'.format(self.TEMP_KEY, self.temp)
+        return str
+
+    def load(self, json_data):
+        if (self.RATE_KEY in json_data):
+            self.rate = json_data[self.RATE_KEY]
+
+        if (self.INIT_ACTION_RATE_KEY in json_data):
+            self.init_action_rate = json_data[self.INIT_ACTION_RATE_KEY]
+
+        if (self.NOISE_KEY in json_data):
+            self.noise = json_data[self.NOISE_KEY]
+
+        if (self.NOISE_INTERNAL_KEY in json_data):
+            self.noise_internal = json_data[self.NOISE_INTERNAL_KEY]
+
+        if (self.TEMP_KEY in json_data):
+            self.temp = json_data[self.TEMP_KEY]
+
+        return
+
+    def lerp(self, other, t):
+        lerp_params = ExpParams()
+        lerp_params.rate = MathUtil.lerp(self.rate, other.rate, t)
+        lerp_params.init_action_rate = MathUtil.lerp(self.init_action_rate, other.init_action_rate, t)
+        lerp_params.noise = MathUtil.lerp(self.noise, other.noise, t)
+        lerp_params.noise_internal = MathUtil.lerp(self.noise_internal, other.noise_internal, t)
+        lerp_params.temp = MathUtil.log_lerp(self.temp, other.temp, t)
+        return lerp_params

examples/pybullet/gym/pybullet_envs/deep_mimic/learning/nets/__init__.py

@@ -0,0 +1 @@
+from . import *

examples/pybullet/gym/pybullet_envs/deep_mimic/learning/nets/fc_2layers_1024units.py

@@ -0,0 +1,13 @@
+import tensorflow as tf
+import learning.tf_util as TFUtil
+
+NAME = "fc_2layers_1024units"
+
+def build_net(input_tfs, reuse=False):
+    layers = [1024, 512]
+    activation = tf.nn.relu
+
+    input_tf = tf.concat(axis=-1, values=input_tfs)          
+    h = TFUtil.fc_net(input_tf, layers, activation=activation, reuse=reuse)
+    h = activation(h)
+    return h

examples/pybullet/gym/pybullet_envs/deep_mimic/learning/nets/net_builder.py

@@ -0,0 +1,11 @@
+import learning.nets.fc_2layers_1024units as fc_2layers_1024units
+
+def build_net(net_name, input_tfs, reuse=False):
+    net = None
+
+    if (net_name == fc_2layers_1024units.NAME):
+        net = fc_2layers_1024units.build_net(input_tfs, reuse)
+    else:
+        assert False, 'Unsupported net: ' + net_name
+    
+    return net

examples/pybullet/gym/pybullet_envs/deep_mimic/learning/normalizer.py

@@ -0,0 +1,149 @@
+import numpy as np
+import copy
+import pybullet_utils.mpi_util as MPIUtil
+from pybullet_utils.logger import Logger
+
+class Normalizer(object):
+    CHECK_SYNC_COUNT = 50000 # check synchronization after a certain number of entries
+
+    # these group IDs must be the same as those in CharController.h
+    NORM_GROUP_SINGLE = 0
+    NORM_GROUP_NONE = -1
+
+    class Group(object):
+        def __init__(self, id, indices):
+            self.id = id
+            self.indices = indices
+            return
+
+    def __init__(self, size, groups_ids=None, eps=0.02, clip=np.inf):
+        self.eps = eps
+        self.clip = clip
+        self.mean = np.zeros(size)
+        self.mean_sq = np.zeros(size)
+        self.std = np.ones(size)
+        self.count = 0
+        self.groups = self._build_groups(groups_ids)
+
+        self.new_count = 0
+        self.new_sum = np.zeros_like(self.mean)
+        self.new_sum_sq = np.zeros_like(self.mean_sq)
+        return
+
+    def record(self, x):
+        size = self.get_size()
+        is_array = isinstance(x, np.ndarray)
+        if not is_array:
+            assert(size == 1)
+            x = np.array([[x]])
+
+        assert x.shape[-1] == size, \
+            Logger.print2('Normalizer shape mismatch, expecting size {:d}, but got {:d}'.format(size, x.shape[-1]))
+        x = np.reshape(x, [-1, size])
+
+        self.new_count += x.shape[0]
+        self.new_sum += np.sum(x, axis=0)
+        self.new_sum_sq += np.sum(np.square(x), axis=0)
+        return
+
+    def update(self):
+        new_count = MPIUtil.reduce_sum(self.new_count)
+        new_sum = MPIUtil.reduce_sum(self.new_sum)
+        new_sum_sq = MPIUtil.reduce_sum(self.new_sum_sq)
+
+        new_total = self.count + new_count
+        if (self.count // self.CHECK_SYNC_COUNT != new_total // self.CHECK_SYNC_COUNT):
+            assert self.check_synced(), Logger.print2('Normalizer parameters desynchronized')
+
+        if new_count > 0:
+            new_mean = self._process_group_data(new_sum / new_count, self.mean)
+            new_mean_sq = self._process_group_data(new_sum_sq / new_count, self.mean_sq)
+            w_old = float(self.count) / new_total
+            w_new = float(new_count) / new_total
+
+            self.mean = w_old * self.mean + w_new * new_mean
+            self.mean_sq = w_old * self.mean_sq + w_new * new_mean_sq
+            self.count = new_total
+            self.std = self.calc_std(self.mean, self.mean_sq)
+
+            self.new_count = 0
+            self.new_sum.fill(0)
+            self.new_sum_sq.fill(0)
+
+        return
+
+    def get_size(self):
+        return self.mean.size
+
+    def set_mean_std(self, mean, std):
+        size = self.get_size()
+        is_array = isinstance(mean, np.ndarray) and isinstance(std, np.ndarray)
+        
+        if not is_array:
+            assert(size == 1)
+            mean = np.array([mean])
+            std = np.array([std])
+
+        assert len(mean) == size and len(std) == size, \
+            Logger.print2('Normalizer shape mismatch, expecting size {:d}, but got {:d} and {:d}'.format(size, len(mean), len(std)))
+        
+        self.mean = mean
+        self.std = std
+        self.mean_sq = self.calc_mean_sq(self.mean, self.std)
+        return
+
+    def normalize(self, x):
+        norm_x = (x - self.mean) / self.std
+        norm_x = np.clip(norm_x, -self.clip, self.clip)
+        return norm_x
+
+    def unnormalize(self, norm_x):
+        x = norm_x * self.std + self.mean
+        return x
+
+    def calc_std(self, mean, mean_sq):
+        var = mean_sq - np.square(mean)
+        # some time floating point errors can lead to small negative numbers
+        var = np.maximum(var, 0)
+        std = np.sqrt(var)
+        std = np.maximum(std, self.eps)
+        return std
+
+    def calc_mean_sq(self, mean, std):
+        return np.square(std) + np.square(self.mean)
+
+    def check_synced(self):
+        synced = True
+        if MPIUtil.is_root_proc():
+            vars = np.concatenate([self.mean, self.mean_sq])
+            MPIUtil.bcast(vars)
+        else:
+            vars_local = np.concatenate([self.mean, self.mean_sq])
+            vars_root = np.empty_like(vars_local)
+            MPIUtil.bcast(vars_root)
+            synced = (vars_local == vars_root).all()
+        return synced
+
+    def _build_groups(self, groups_ids):
+        groups = []
+        if groups_ids is None:
+            curr_id = self.NORM_GROUP_SINGLE
+            curr_list = np.arange(self.get_size()).astype(np.int32)
+            groups.append(self.Group(curr_id, curr_list))
+        else:
+            ids = np.unique(groups_ids)
+            for id in ids:
+                curr_list = np.nonzero(groups_ids == id)[0].astype(np.int32)
+                groups.append(self.Group(id, curr_list))
+
+        return groups
+
+    def _process_group_data(self, new_data, old_data):
+        proc_data = new_data.copy()
+        for group in self.groups:
+            if group.id == self.NORM_GROUP_NONE:
+                proc_data[group.indices] = old_data[group.indices]
+            elif group.id != self.NORM_GROUP_SINGLE:
+                avg = np.mean(new_data[group.indices])
+                proc_data[group.indices] = avg
+        return proc_data

examples/pybullet/gym/pybullet_envs/deep_mimic/learning/path.py

@@ -0,0 +1,46 @@
+import numpy as np
+from pybullet_envs.deep_mimic.env.env import Env
+
+class Path(object):
+    def __init__(self):
+        self.clear()
+        return
+
+    def pathlength(self):
+        return len(self.actions)
+
+    def is_valid(self):
+        valid = True
+        l = self.pathlength()
+        valid &= len(self.states) == l + 1
+        valid &= len(self.goals) == l + 1
+        valid &= len(self.actions) == l
+        valid &= len(self.logps) == l
+        valid &= len(self.rewards) == l
+        valid &= len(self.flags) == l
+
+        return valid
+
+    def check_vals(self):
+        for vals in [self.states, self.goals, self.actions, self.logps,
+                  self.rewards]:
+            for v in vals:
+                if not np.isfinite(v).all():
+                    return False
+        return True
+
+    def clear(self):
+        self.states = []
+        self.goals = []
+        self.actions = []
+        self.logps = []
+        self.rewards = []
+        self.flags = []
+        self.terminate = Env.Terminate.Null
+        return
+
+    def get_pathlen(self):
+        return len(self.rewards)
+
+    def calc_return(self):
+        return sum(self.rewards)

examples/pybullet/gym/pybullet_envs/deep_mimic/learning/pg_agent.py

@@ -0,0 +1,353 @@
+import numpy as np
+import tensorflow as tf
+import copy
+
+from learning.tf_agent import TFAgent
+from learning.solvers.mpi_solver import MPISolver
+import learning.tf_util as TFUtil
+import learning.nets.net_builder as NetBuilder
+from learning.tf_normalizer import TFNormalizer
+import learning.rl_util as RLUtil
+from pybullet_utils.logger import Logger
+import pybullet_utils.mpi_util as MPIUtil
+import pybullet_utils.math_util as MathUtil
+from pybullet_envs.deep_mimic.env.action_space import ActionSpace
+from pybullet_envs.deep_mimic.env.env import Env
+
+'''
+Policy Gradient Agent
+'''
+
+class PGAgent(TFAgent):
+    NAME = 'PG'
+
+    ACTOR_NET_KEY = 'ActorNet'
+    ACTOR_STEPSIZE_KEY = 'ActorStepsize'
+    ACTOR_MOMENTUM_KEY = 'ActorMomentum'
+    ACTOR_WEIGHT_DECAY_KEY = 'ActorWeightDecay'
+    ACTOR_INIT_OUTPUT_SCALE_KEY = 'ActorInitOutputScale'
+
+    CRITIC_NET_KEY = 'CriticNet'
+    CRITIC_STEPSIZE_KEY = 'CriticStepsize'
+    CRITIC_MOMENTUM_KEY = 'CriticMomentum'
+    CRITIC_WEIGHT_DECAY_KEY = 'CriticWeightDecay'
+    
+    EXP_ACTION_FLAG = 1 << 0
+
+    def __init__(self, world, id, json_data): 
+        self._exp_action = False
+        super().__init__(world, id, json_data)
+        return
+
+    def reset(self):
+        super().reset()
+        self._exp_action = False
+        return
+
+    def _check_action_space(self):
+        action_space = self.get_action_space()
+        return action_space == ActionSpace.Continuous
+
+    def _load_params(self, json_data):
+        super()._load_params(json_data)
+        self.val_min, self.val_max = self._calc_val_bounds(self.discount)
+        self.val_fail, self.val_succ = self._calc_term_vals(self.discount)
+        return
+
+    def _build_nets(self, json_data):
+        assert self.ACTOR_NET_KEY in json_data
+        assert self.CRITIC_NET_KEY in json_data
+
+        actor_net_name = json_data[self.ACTOR_NET_KEY]
+        critic_net_name = json_data[self.CRITIC_NET_KEY]
+        actor_init_output_scale = 1 if (self.ACTOR_INIT_OUTPUT_SCALE_KEY not in json_data) else json_data[self.ACTOR_INIT_OUTPUT_SCALE_KEY]
+        
+        s_size = self.get_state_size()
+        g_size = self.get_goal_size()
+        a_size = self.get_action_size()
+
+        # setup input tensors
+        self.s_tf = tf.placeholder(tf.float32, shape=[None, s_size], name="s") # observations
+        self.tar_val_tf = tf.placeholder(tf.float32, shape=[None], name="tar_val") # target value s
+        self.adv_tf = tf.placeholder(tf.float32, shape=[None], name="adv") # advantage
+        self.a_tf = tf.placeholder(tf.float32, shape=[None, a_size], name="a") # target actions
+        self.g_tf = tf.placeholder(tf.float32, shape=([None, g_size] if self.has_goal() else None), name="g") # goals
+
+        with tf.variable_scope('main'):
+            with tf.variable_scope('actor'):
+                self.actor_tf = self._build_net_actor(actor_net_name, actor_init_output_scale)
+            with tf.variable_scope('critic'):
+                self.critic_tf = self._build_net_critic(critic_net_name)
+
+        if (self.actor_tf != None):
+            Logger.print2('Built actor net: ' + actor_net_name)
+
+        if (self.critic_tf != None):
+            Logger.print2('Built critic net: ' + critic_net_name)
+
+        return
+
+    def _build_normalizers(self):
+        super()._build_normalizers()
+        with self.sess.as_default(), self.graph.as_default(), tf.variable_scope(self.tf_scope):
+            with tf.variable_scope(self.RESOURCE_SCOPE):
+                val_offset, val_scale = self._calc_val_offset_scale(self.discount)
+                self.val_norm = TFNormalizer(self.sess, 'val_norm', 1)
+                self.val_norm.set_mean_std(-val_offset, 1.0 / val_scale)
+        return
+
+    def _init_normalizers(self):
+        super()._init_normalizers()
+        with self.sess.as_default(), self.graph.as_default():
+            self.val_norm.update()
+        return
+
+    def _load_normalizers(self):
+        super()._load_normalizers()
+        self.val_norm.load()
+        return
+
+    def _build_losses(self, json_data):
+        actor_weight_decay = 0 if (self.ACTOR_WEIGHT_DECAY_KEY not in json_data) else json_data[self.ACTOR_WEIGHT_DECAY_KEY]
+        critic_weight_decay = 0 if (self.CRITIC_WEIGHT_DECAY_KEY not in json_data) else json_data[self.CRITIC_WEIGHT_DECAY_KEY]
+
+        norm_val_diff = self.val_norm.normalize_tf(self.tar_val_tf) - self.val_norm.normalize_tf(self.critic_tf)
+        self.critic_loss_tf = 0.5 * tf.reduce_mean(tf.square(norm_val_diff))
+
+        if (critic_weight_decay != 0):
+            self.critic_loss_tf += critic_weight_decay * self._weight_decay_loss('main/critic')
+        
+        norm_a_mean_tf = self.a_norm.normalize_tf(self.actor_tf)
+        norm_a_diff = self.a_norm.normalize_tf(self.a_tf) - norm_a_mean_tf
+
+        self.actor_loss_tf = tf.reduce_sum(tf.square(norm_a_diff), axis=-1)
+        self.actor_loss_tf *= self.adv_tf
+        self.actor_loss_tf = 0.5 * tf.reduce_mean(self.actor_loss_tf)
+
+        norm_a_bound_min = self.a_norm.normalize(self.a_bound_min)
+        norm_a_bound_max = self.a_norm.normalize(self.a_bound_max)
+        a_bound_loss = TFUtil.calc_bound_loss(norm_a_mean_tf, norm_a_bound_min, norm_a_bound_max)
+        a_bound_loss /= self.exp_params_curr.noise
+        self.actor_loss_tf += a_bound_loss
+
+        if (actor_weight_decay != 0):
+            self.actor_loss_tf += actor_weight_decay * self._weight_decay_loss('main/actor')
+        
+        return
+
+    def _build_solvers(self, json_data):
+        actor_stepsize = 0.001 if (self.ACTOR_STEPSIZE_KEY not in json_data) else json_data[self.ACTOR_STEPSIZE_KEY]
+        actor_momentum = 0.9 if (self.ACTOR_MOMENTUM_KEY not in json_data) else json_data[self.ACTOR_MOMENTUM_KEY]
+        critic_stepsize = 0.01 if (self.CRITIC_STEPSIZE_KEY not in json_data) else json_data[self.CRITIC_STEPSIZE_KEY]
+        critic_momentum = 0.9 if (self.CRITIC_MOMENTUM_KEY not in json_data) else json_data[self.CRITIC_MOMENTUM_KEY]
+        
+        critic_vars = self._tf_vars('main/critic')
+        critic_opt = tf.train.MomentumOptimizer(learning_rate=critic_stepsize, momentum=critic_momentum)
+        self.critic_grad_tf = tf.gradients(self.critic_loss_tf, critic_vars)
+        self.critic_solver = MPISolver(self.sess, critic_opt, critic_vars)
+
+        actor_vars = self._tf_vars('main/actor')
+        actor_opt = tf.train.MomentumOptimizer(learning_rate=actor_stepsize, momentum=actor_momentum)
+        self.actor_grad_tf = tf.gradients(self.actor_loss_tf, actor_vars)
+        self.actor_solver = MPISolver(self.sess, actor_opt, actor_vars)
+
+        return
+
+    def _build_net_actor(self, net_name, init_output_scale):
+        norm_s_tf = self.s_norm.normalize_tf(self.s_tf)
+        input_tfs = [norm_s_tf]
+        if (self.has_goal()):
+            norm_g_tf = self.g_norm.normalize_tf(self.g_tf)
+            input_tfs += [norm_g_tf]
+        
+        h = NetBuilder.build_net(net_name, input_tfs)
+        norm_a_tf = tf.layers.dense(inputs=h, units=self.get_action_size(), activation=None,
+                                kernel_initializer=tf.random_uniform_initializer(minval=-init_output_scale, maxval=init_output_scale))
+        
+        a_tf = self.a_norm.unnormalize_tf(norm_a_tf)
+        return a_tf
+    
+    def _build_net_critic(self, net_name):
+        norm_s_tf = self.s_norm.normalize_tf(self.s_tf)
+        input_tfs = [norm_s_tf]
+        if (self.has_goal()):
+            norm_g_tf = self.g_norm.normalize_tf(self.g_tf)
+            input_tfs += [norm_g_tf]
+        
+        h = NetBuilder.build_net(net_name, input_tfs)
+        norm_val_tf = tf.layers.dense(inputs=h, units=1, activation=None,
+                                kernel_initializer=TFUtil.xavier_initializer);
+
+        norm_val_tf = tf.reshape(norm_val_tf, [-1])
+        val_tf = self.val_norm.unnormalize_tf(norm_val_tf)
+        return val_tf
+
+    def _initialize_vars(self):
+        super()._initialize_vars()
+        self._sync_solvers()
+        return
+
+    def _sync_solvers(self):
+        self.actor_solver.sync()
+        self.critic_solver.sync()
+        return
+
+    def _decide_action(self, s, g):
+        with self.sess.as_default(), self.graph.as_default():
+            self._exp_action = False
+            a = self._eval_actor(s, g)[0]
+            logp = 0
+
+            if self._enable_stoch_policy():
+                # epsilon-greedy
+                rand_action = MathUtil.flip_coin(self.exp_params_curr.rate)
+                if rand_action:
+                    norm_exp_noise = np.random.randn(*a.shape)
+                    norm_exp_noise *= self.exp_params_curr.noise
+                    exp_noise = norm_exp_noise * self.a_norm.std
+                    a += exp_noise
+
+                    logp = self._calc_action_logp(norm_exp_noise)
+                    self._exp_action = True
+
+        return a, logp
+
+    def _enable_stoch_policy(self):
+        return self.enable_training and (self._mode == self.Mode.TRAIN or self._mode == self.Mode.TRAIN_END)
+
+    def _eval_actor(self, s, g):
+        s = np.reshape(s, [-1, self.get_state_size()])
+        g = np.reshape(g, [-1, self.get_goal_size()]) if self.has_goal() else None
+          
+        feed = {
+            self.s_tf : s,
+            self.g_tf : g
+        }
+
+        a = self.actor_tf.eval(feed)
+        return a
+    
+    def _eval_critic(self, s, g):
+        with self.sess.as_default(), self.graph.as_default():
+            s = np.reshape(s, [-1, self.get_state_size()])
+            g = np.reshape(g, [-1, self.get_goal_size()]) if self.has_goal() else None
+
+            feed = {
+                self.s_tf : s,
+                self.g_tf : g
+            }
+
+            val = self.critic_tf.eval(feed)    
+        return val
+
+    def _record_flags(self):
+        flags = int(0)
+        if (self._exp_action):
+            flags = flags | self.EXP_ACTION_FLAG
+        return flags
+
+    def _train_step(self):
+        super()._train_step()
+
+        critic_loss = self._update_critic()
+        actor_loss = self._update_actor()
+        critic_loss = MPIUtil.reduce_avg(critic_loss)
+        actor_loss = MPIUtil.reduce_avg(actor_loss)
+
+        critic_stepsize = self.critic_solver.get_stepsize()
+        actor_stepsize = self.actor_solver.get_stepsize()
+        
+        self.logger.log_tabular('Critic_Loss', critic_loss)
+        self.logger.log_tabular('Critic_Stepsize', critic_stepsize)
+        self.logger.log_tabular('Actor_Loss', actor_loss) 
+        self.logger.log_tabular('Actor_Stepsize', actor_stepsize)
+
+        return
+
+    def _update_critic(self):
+        idx = self.replay_buffer.sample(self._local_mini_batch_size)
+        s = self.replay_buffer.get('states', idx)
+        g = self.replay_buffer.get('goals', idx) if self.has_goal() else None
+        
+        tar_V = self._calc_updated_vals(idx)
+        tar_V = np.clip(tar_V, self.val_min, self.val_max)
+
+        feed = {
+            self.s_tf: s,
+            self.g_tf: g,
+            self.tar_val_tf: tar_V
+        }
+
+        loss, grads = self.sess.run([self.critic_loss_tf, self.critic_grad_tf], feed)
+        self.critic_solver.update(grads)
+        return loss
+    
+    def _update_actor(self):
+        key = self.EXP_ACTION_FLAG
+        idx = self.replay_buffer.sample_filtered(self._local_mini_batch_size, key)
+        has_goal = self.has_goal()
+
+        s = self.replay_buffer.get('states', idx)
+        g = self.replay_buffer.get('goals', idx) if has_goal else None
+        a = self.replay_buffer.get('actions', idx)
+
+        V_new = self._calc_updated_vals(idx)
+        V_old = self._eval_critic(s, g)
+        adv = V_new - V_old
+
+        feed = {
+            self.s_tf: s,
+            self.g_tf: g,
+            self.a_tf: a,
+            self.adv_tf: adv
+        }
+
+        loss, grads = self.sess.run([self.actor_loss_tf, self.actor_grad_tf], feed)
+        self.actor_solver.update(grads)
+
+        return loss
+
+    def _calc_updated_vals(self, idx):
+        r = self.replay_buffer.get('rewards', idx)
+
+        if self.discount == 0:
+            new_V = r
+        else:
+            next_idx = self.replay_buffer.get_next_idx(idx)
+            s_next = self.replay_buffer.get('states', next_idx)
+            g_next = self.replay_buffer.get('goals', next_idx) if self.has_goal() else None
+
+            is_end = self.replay_buffer.is_path_end(idx)
+            is_fail = self.replay_buffer.check_terminal_flag(idx, Env.Terminate.Fail)
+            is_succ = self.replay_buffer.check_terminal_flag(idx, Env.Terminate.Succ)
+            is_fail = np.logical_and(is_end, is_fail) 
+            is_succ = np.logical_and(is_end, is_succ) 
+
+            V_next = self._eval_critic(s_next, g_next)
+            V_next[is_fail] = self.val_fail
+            V_next[is_succ] = self.val_succ
+
+            new_V = r + self.discount * V_next
+        return new_V
+
+    def _calc_action_logp(self, norm_action_deltas):
+        # norm action delta are for the normalized actions (scaled by self.a_norm.std) 
+        stdev = self.exp_params_curr.noise
+        assert stdev > 0
+
+        a_size = self.get_action_size()
+        logp = -0.5 / (stdev * stdev) * np.sum(np.square(norm_action_deltas), axis=-1)
+        logp += -0.5 * a_size * np.log(2 * np.pi)
+        logp += -a_size * np.log(stdev)
+        return logp
+
+    def _log_val(self, s, g):
+        val = self._eval_critic(s, g)
+        norm_val = self.val_norm.normalize(val)
+        self.world.env.log_val(self.id, norm_val[0])
+        return
+
+    def _build_replay_buffer(self, buffer_size):
+        super()._build_replay_buffer(buffer_size)
+        self.replay_buffer.add_filter_key(self.EXP_ACTION_FLAG)
+        return

examples/pybullet/gym/pybullet_envs/deep_mimic/learning/ppo_agent.py

@@ -0,0 +1,367 @@
+import numpy as np
+import copy as copy
+import tensorflow as tf
+
+from learning.pg_agent import PGAgent
+from learning.solvers.mpi_solver import MPISolver
+import learning.tf_util as TFUtil
+import learning.rl_util as RLUtil
+from pybullet_utils.logger import Logger
+import pybullet_utils.mpi_util as MPIUtil
+import pybullet_utils.math_util as MathUtil
+from pybullet_envs.deep_mimic.env.env import Env
+
+'''
+Proximal Policy Optimization Agent
+'''
+
+class PPOAgent(PGAgent):
+    NAME = "PPO"
+    EPOCHS_KEY = "Epochs"
+    BATCH_SIZE_KEY = "BatchSize"
+    RATIO_CLIP_KEY = "RatioClip"
+    NORM_ADV_CLIP_KEY = "NormAdvClip"
+    TD_LAMBDA_KEY = "TDLambda"
+    TAR_CLIP_FRAC = "TarClipFrac"
+    ACTOR_STEPSIZE_DECAY = "ActorStepsizeDecay"
+
+    def __init__(self, world, id, json_data): 
+        super().__init__(world, id, json_data)
+        return
+
+    def _load_params(self, json_data):
+        super()._load_params(json_data)
+
+        self.epochs = 1 if (self.EPOCHS_KEY not in json_data) else json_data[self.EPOCHS_KEY]
+        self.batch_size = 1024 if (self.BATCH_SIZE_KEY not in json_data) else json_data[self.BATCH_SIZE_KEY]
+        self.ratio_clip = 0.2 if (self.RATIO_CLIP_KEY not in json_data) else json_data[self.RATIO_CLIP_KEY]
+        self.norm_adv_clip = 5 if (self.NORM_ADV_CLIP_KEY not in json_data) else json_data[self.NORM_ADV_CLIP_KEY]
+        self.td_lambda = 0.95 if (self.TD_LAMBDA_KEY not in json_data) else json_data[self.TD_LAMBDA_KEY]
+        self.tar_clip_frac = -1 if (self.TAR_CLIP_FRAC not in json_data) else json_data[self.TAR_CLIP_FRAC]
+        self.actor_stepsize_decay = 0.5 if (self.ACTOR_STEPSIZE_DECAY not in json_data) else json_data[self.ACTOR_STEPSIZE_DECAY]
+
+        num_procs = MPIUtil.get_num_procs()
+        local_batch_size = int(self.batch_size / num_procs)
+        min_replay_size = 2 * local_batch_size # needed to prevent buffer overflow
+        assert(self.replay_buffer_size > min_replay_size)
+
+        self.replay_buffer_size = np.maximum(min_replay_size, self.replay_buffer_size)
+
+        return
+
+    def _build_nets(self, json_data):
+        assert self.ACTOR_NET_KEY in json_data
+        assert self.CRITIC_NET_KEY in json_data
+
+        actor_net_name = json_data[self.ACTOR_NET_KEY]
+        critic_net_name = json_data[self.CRITIC_NET_KEY]
+        actor_init_output_scale = 1 if (self.ACTOR_INIT_OUTPUT_SCALE_KEY not in json_data) else json_data[self.ACTOR_INIT_OUTPUT_SCALE_KEY]
+
+        s_size = self.get_state_size()
+        g_size = self.get_goal_size()
+        a_size = self.get_action_size()
+
+        # setup input tensors
+        self.s_tf = tf.placeholder(tf.float32, shape=[None, s_size], name="s")
+        self.a_tf = tf.placeholder(tf.float32, shape=[None, a_size], name="a")
+        self.tar_val_tf = tf.placeholder(tf.float32, shape=[None], name="tar_val")
+        self.adv_tf = tf.placeholder(tf.float32, shape=[None], name="adv")
+        self.g_tf = tf.placeholder(tf.float32, shape=([None, g_size] if self.has_goal() else None), name="g")
+        self.old_logp_tf = tf.placeholder(tf.float32, shape=[None], name="old_logp")
+        self.exp_mask_tf = tf.placeholder(tf.float32, shape=[None], name="exp_mask")
+
+        with tf.variable_scope('main'):
+            with tf.variable_scope('actor'):
+                self.a_mean_tf = self._build_net_actor(actor_net_name, actor_init_output_scale)
+            with tf.variable_scope('critic'):
+                self.critic_tf = self._build_net_critic(critic_net_name)
+                
+        if (self.a_mean_tf != None):
+            Logger.print2('Built actor net: ' + actor_net_name)
+
+        if (self.critic_tf != None):
+            Logger.print2('Built critic net: ' + critic_net_name)
+        
+        self.norm_a_std_tf = self.exp_params_curr.noise * tf.ones(a_size)
+        norm_a_noise_tf = self.norm_a_std_tf * tf.random_normal(shape=tf.shape(self.a_mean_tf))
+        norm_a_noise_tf *= tf.expand_dims(self.exp_mask_tf, axis=-1)
+        self.sample_a_tf = self.a_mean_tf + norm_a_noise_tf * self.a_norm.std_tf
+        self.sample_a_logp_tf = TFUtil.calc_logp_gaussian(x_tf=norm_a_noise_tf, mean_tf=None, std_tf=self.norm_a_std_tf)
+
+        return
+
+    def _build_losses(self, json_data):
+        actor_weight_decay = 0 if (self.ACTOR_WEIGHT_DECAY_KEY not in json_data) else json_data[self.ACTOR_WEIGHT_DECAY_KEY]
+        critic_weight_decay = 0 if (self.CRITIC_WEIGHT_DECAY_KEY not in json_data) else json_data[self.CRITIC_WEIGHT_DECAY_KEY]
+        
+        norm_val_diff = self.val_norm.normalize_tf(self.tar_val_tf) - self.val_norm.normalize_tf(self.critic_tf)
+        self.critic_loss_tf = 0.5 * tf.reduce_mean(tf.square(norm_val_diff))
+
+        if (critic_weight_decay != 0):
+            self.critic_loss_tf += critic_weight_decay * self._weight_decay_loss('main/critic')
+        
+        norm_tar_a_tf = self.a_norm.normalize_tf(self.a_tf)
+        self._norm_a_mean_tf = self.a_norm.normalize_tf(self.a_mean_tf)
+
+        self.logp_tf = TFUtil.calc_logp_gaussian(norm_tar_a_tf, self._norm_a_mean_tf, self.norm_a_std_tf)
+        ratio_tf = tf.exp(self.logp_tf - self.old_logp_tf)
+        actor_loss0 = self.adv_tf * ratio_tf
+        actor_loss1 = self.adv_tf * tf.clip_by_value(ratio_tf, 1.0 - self.ratio_clip, 1 + self.ratio_clip)
+        self.actor_loss_tf = -tf.reduce_mean(tf.minimum(actor_loss0, actor_loss1))
+
+        norm_a_bound_min = self.a_norm.normalize(self.a_bound_min)
+        norm_a_bound_max = self.a_norm.normalize(self.a_bound_max)
+        a_bound_loss = TFUtil.calc_bound_loss(self._norm_a_mean_tf, norm_a_bound_min, norm_a_bound_max)
+        self.actor_loss_tf += a_bound_loss
+
+        if (actor_weight_decay != 0):
+            self.actor_loss_tf += actor_weight_decay * self._weight_decay_loss('main/actor')
+        
+        # for debugging
+        self.clip_frac_tf = tf.reduce_mean(tf.to_float(tf.greater(tf.abs(ratio_tf - 1.0), self.ratio_clip)))
+
+        return
+
+    def _build_solvers(self, json_data):
+        actor_stepsize = 0.001 if (self.ACTOR_STEPSIZE_KEY not in json_data) else json_data[self.ACTOR_STEPSIZE_KEY]
+        actor_momentum = 0.9 if (self.ACTOR_MOMENTUM_KEY not in json_data) else json_data[self.ACTOR_MOMENTUM_KEY]
+        critic_stepsize = 0.01 if (self.CRITIC_STEPSIZE_KEY not in json_data) else json_data[self.CRITIC_STEPSIZE_KEY]
+        critic_momentum = 0.9 if (self.CRITIC_MOMENTUM_KEY not in json_data) else json_data[self.CRITIC_MOMENTUM_KEY]
+        
+        critic_vars = self._tf_vars('main/critic')
+        critic_opt = tf.train.MomentumOptimizer(learning_rate=critic_stepsize, momentum=critic_momentum)
+        self.critic_grad_tf = tf.gradients(self.critic_loss_tf, critic_vars)
+        self.critic_solver = MPISolver(self.sess, critic_opt, critic_vars)
+
+        self._actor_stepsize_tf = tf.get_variable(dtype=tf.float32, name='actor_stepsize', initializer=actor_stepsize, trainable=False)
+        self._actor_stepsize_ph = tf.get_variable(dtype=tf.float32, name='actor_stepsize_ph', shape=[])
+        self._actor_stepsize_update_op = self._actor_stepsize_tf.assign(self._actor_stepsize_ph)
+
+        actor_vars = self._tf_vars('main/actor')
+        actor_opt = tf.train.MomentumOptimizer(learning_rate=self._actor_stepsize_tf, momentum=actor_momentum)
+        self.actor_grad_tf = tf.gradients(self.actor_loss_tf, actor_vars)
+        self.actor_solver = MPISolver(self.sess, actor_opt, actor_vars)
+        
+        return
+
+    def _decide_action(self, s, g):
+        with self.sess.as_default(), self.graph.as_default():
+            self._exp_action = self._enable_stoch_policy() and MathUtil.flip_coin(self.exp_params_curr.rate)
+            a, logp = self._eval_actor(s, g, self._exp_action)
+        return a[0], logp[0]
+
+    def _eval_actor(self, s, g, enable_exp):
+        s = np.reshape(s, [-1, self.get_state_size()])
+        g = np.reshape(g, [-1, self.get_goal_size()]) if self.has_goal() else None
+          
+        feed = {
+            self.s_tf : s,
+            self.g_tf : g,
+            self.exp_mask_tf: np.array([1 if enable_exp else 0])
+        }
+
+        a, logp = self.sess.run([self.sample_a_tf, self.sample_a_logp_tf], feed_dict=feed)
+        return a, logp
+
+    def _train_step(self):
+        adv_eps = 1e-5
+
+        start_idx = self.replay_buffer.buffer_tail
+        end_idx = self.replay_buffer.buffer_head
+        assert(start_idx == 0)
+        assert(self.replay_buffer.get_current_size() <= self.replay_buffer.buffer_size) # must avoid overflow
+        assert(start_idx < end_idx)
+
+        idx = np.array(list(range(start_idx, end_idx)))        
+        end_mask = self.replay_buffer.is_path_end(idx)
+        end_mask = np.logical_not(end_mask) 
+        
+        vals = self._compute_batch_vals(start_idx, end_idx)
+        new_vals = self._compute_batch_new_vals(start_idx, end_idx, vals)
+
+        valid_idx = idx[end_mask]
+        exp_idx = self.replay_buffer.get_idx_filtered(self.EXP_ACTION_FLAG).copy()
+        num_valid_idx = valid_idx.shape[0]
+        num_exp_idx = exp_idx.shape[0]
+        exp_idx = np.column_stack([exp_idx, np.array(list(range(0, num_exp_idx)), dtype=np.int32)])
+        
+        local_sample_count = valid_idx.size
+        global_sample_count = int(MPIUtil.reduce_sum(local_sample_count))
+        mini_batches = int(np.ceil(global_sample_count / self.mini_batch_size))
+        
+        adv = new_vals[exp_idx[:,0]] - vals[exp_idx[:,0]]
+        new_vals = np.clip(new_vals, self.val_min, self.val_max)
+
+        adv_mean = np.mean(adv)
+        adv_std = np.std(adv)
+        adv = (adv - adv_mean) / (adv_std + adv_eps)
+        adv = np.clip(adv, -self.norm_adv_clip, self.norm_adv_clip)
+
+        critic_loss = 0
+        actor_loss = 0
+        actor_clip_frac = 0
+
+        for e in range(self.epochs):
+            np.random.shuffle(valid_idx)
+            np.random.shuffle(exp_idx)
+
+            for b in range(mini_batches):
+                batch_idx_beg = b * self._local_mini_batch_size
+                batch_idx_end = batch_idx_beg + self._local_mini_batch_size
+
+                critic_batch = np.array(range(batch_idx_beg, batch_idx_end), dtype=np.int32)
+                actor_batch = critic_batch.copy()
+                critic_batch = np.mod(critic_batch, num_valid_idx)
+                actor_batch = np.mod(actor_batch, num_exp_idx)
+                shuffle_actor = (actor_batch[-1] < actor_batch[0]) or (actor_batch[-1] == num_exp_idx - 1)
+
+                critic_batch = valid_idx[critic_batch]
+                actor_batch = exp_idx[actor_batch]
+                critic_batch_vals = new_vals[critic_batch]
+                actor_batch_adv = adv[actor_batch[:,1]]
+
+                critic_s = self.replay_buffer.get('states', critic_batch)
+                critic_g = self.replay_buffer.get('goals', critic_batch) if self.has_goal() else None
+                curr_critic_loss = self._update_critic(critic_s, critic_g, critic_batch_vals)
+
+                actor_s = self.replay_buffer.get("states", actor_batch[:,0])
+                actor_g = self.replay_buffer.get("goals", actor_batch[:,0]) if self.has_goal() else None
+                actor_a = self.replay_buffer.get("actions", actor_batch[:,0])
+                actor_logp = self.replay_buffer.get("logps", actor_batch[:,0])
+                curr_actor_loss, curr_actor_clip_frac = self._update_actor(actor_s, actor_g, actor_a, actor_logp, actor_batch_adv)
+                
+                critic_loss += curr_critic_loss
+                actor_loss += np.abs(curr_actor_loss)
+                actor_clip_frac += curr_actor_clip_frac
+
+                if (shuffle_actor):
+                    np.random.shuffle(exp_idx)
+
+        total_batches = mini_batches * self.epochs
+        critic_loss /= total_batches
+        actor_loss /= total_batches
+        actor_clip_frac /= total_batches
+
+        critic_loss = MPIUtil.reduce_avg(critic_loss)
+        actor_loss = MPIUtil.reduce_avg(actor_loss)
+        actor_clip_frac = MPIUtil.reduce_avg(actor_clip_frac)
+
+        critic_stepsize = self.critic_solver.get_stepsize()
+        actor_stepsize = self.update_actor_stepsize(actor_clip_frac)
+
+        self.logger.log_tabular('Critic_Loss', critic_loss)
+        self.logger.log_tabular('Critic_Stepsize', critic_stepsize)
+        self.logger.log_tabular('Actor_Loss', actor_loss) 
+        self.logger.log_tabular('Actor_Stepsize', actor_stepsize)
+        self.logger.log_tabular('Clip_Frac', actor_clip_frac)
+        self.logger.log_tabular('Adv_Mean', adv_mean)
+        self.logger.log_tabular('Adv_Std', adv_std)
+
+        self.replay_buffer.clear()
+
+        return
+
+    def _get_iters_per_update(self):
+        return 1
+
+    def _valid_train_step(self):
+        samples = self.replay_buffer.get_current_size()
+        exp_samples = self.replay_buffer.count_filtered(self.EXP_ACTION_FLAG)
+        global_sample_count = int(MPIUtil.reduce_sum(samples))
+        global_exp_min = int(MPIUtil.reduce_min(exp_samples))
+        return (global_sample_count > self.batch_size) and (global_exp_min > 0)
+
+    def _compute_batch_vals(self, start_idx, end_idx):
+        states = self.replay_buffer.get_all("states")[start_idx:end_idx]
+        goals = self.replay_buffer.get_all("goals")[start_idx:end_idx] if self.has_goal() else None
+        
+        idx = np.array(list(range(start_idx, end_idx)))        
+        is_end = self.replay_buffer.is_path_end(idx)
+        is_fail = self.replay_buffer.check_terminal_flag(idx, Env.Terminate.Fail)
+        is_succ = self.replay_buffer.check_terminal_flag(idx, Env.Terminate.Succ)
+        is_fail = np.logical_and(is_end, is_fail) 
+        is_succ = np.logical_and(is_end, is_succ) 
+
+        vals = self._eval_critic(states, goals)
+        vals[is_fail] = self.val_fail
+        vals[is_succ] = self.val_succ
+
+        return vals
+
+    def _compute_batch_new_vals(self, start_idx, end_idx, val_buffer):
+        rewards = self.replay_buffer.get_all("rewards")[start_idx:end_idx]
+
+        if self.discount == 0:
+            new_vals = rewards.copy()
+        else:
+            new_vals = np.zeros_like(val_buffer)
+
+            curr_idx = start_idx
+            while curr_idx < end_idx:
+                idx0 = curr_idx - start_idx
+                idx1 = self.replay_buffer.get_path_end(curr_idx) - start_idx
+                r = rewards[idx0:idx1]
+                v = val_buffer[idx0:(idx1 + 1)]
+
+                new_vals[idx0:idx1] = RLUtil.compute_return(r, self.discount, self.td_lambda, v)
+                curr_idx = idx1 + start_idx + 1
+        
+        return new_vals
+
+    def _update_critic(self, s, g, tar_vals):
+        feed = {
+            self.s_tf: s,
+            self.g_tf: g,
+            self.tar_val_tf: tar_vals
+        }
+
+        loss, grads = self.sess.run([self.critic_loss_tf, self.critic_grad_tf], feed)
+        self.critic_solver.update(grads)
+        return loss
+    
+    def _update_actor(self, s, g, a, logp, adv):
+        feed = {
+            self.s_tf: s,
+            self.g_tf: g,
+            self.a_tf: a,
+            self.adv_tf: adv,
+            self.old_logp_tf: logp
+        }
+
+        loss, grads, clip_frac = self.sess.run([self.actor_loss_tf, self.actor_grad_tf,
+                                                        self.clip_frac_tf], feed)
+        self.actor_solver.update(grads)
+
+        return loss, clip_frac
+
+    def update_actor_stepsize(self, clip_frac):
+        clip_tol = 1.5
+        step_scale = 2
+        max_stepsize = 1e-2
+        min_stepsize = 1e-8
+        warmup_iters = 5
+
+        actor_stepsize = self.actor_solver.get_stepsize()
+        if (self.tar_clip_frac >= 0 and self.iter > warmup_iters):
+            min_clip = self.tar_clip_frac / clip_tol
+            max_clip = self.tar_clip_frac * clip_tol
+            under_tol = clip_frac < min_clip
+            over_tol = clip_frac > max_clip
+
+            if (over_tol or under_tol):
+                if (over_tol):
+                    actor_stepsize *= self.actor_stepsize_decay
+                else:
+                    actor_stepsize /= self.actor_stepsize_decay
+
+                actor_stepsize = np.clip(actor_stepsize, min_stepsize, max_stepsize)
+                self.set_actor_stepsize(actor_stepsize)
+
+        return actor_stepsize
+
+    def set_actor_stepsize(self, stepsize):
+        feed = {
+            self._actor_stepsize_ph: stepsize,
+        }
+        self.sess.run(self._actor_stepsize_update_op, feed)
+        return

examples/pybullet/gym/pybullet_envs/deep_mimic/learning/replay_buffer.py

@@ -0,0 +1,351 @@
+import numpy as np
+import copy
+from pybullet_utils.logger import Logger
+import inspect as inspect
+from env.env import Env
+import pybullet_utils.math_util as MathUtil
+
+class ReplayBuffer(object):
+    TERMINATE_KEY = 'terminate'
+    PATH_START_KEY = 'path_start'
+    PATH_END_KEY = 'path_end'
+
+    def __init__(self, buffer_size):
+        assert buffer_size > 0
+
+        self.buffer_size = buffer_size
+        self.total_count = 0
+        self.buffer_head = 0
+        self.buffer_tail = MathUtil.INVALID_IDX
+        self.num_paths = 0
+        self._sample_buffers = dict()
+        self.buffers = None
+
+        self.clear()
+        return
+
+    def sample(self, n):
+        curr_size = self.get_current_size()
+        assert curr_size > 0
+
+        idx = np.empty(n, dtype=int)
+        # makes sure that the end states are not sampled
+        for i in range(n):
+            while True:
+                curr_idx = np.random.randint(0, curr_size, size=1)[0]
+                curr_idx += self.buffer_tail
+                curr_idx = np.mod(curr_idx, self.buffer_size)
+
+                if not self.is_path_end(curr_idx):
+                    break
+            idx[i] = curr_idx
+
+        return idx
+
+    def sample_filtered(self, n, key):
+        assert key in self._sample_buffers
+        curr_buffer = self._sample_buffers[key]
+        idx = curr_buffer.sample(n)
+        return idx
+
+    def count_filtered(self, key):
+        curr_buffer = self._sample_buffers[key]
+        return curr_buffer.count
+
+    def get(self, key, idx):
+        return self.buffers[key][idx]
+
+    def get_all(self, key):
+        return self.buffers[key]
+
+    def get_idx_filtered(self, key):
+        assert key in self._sample_buffers
+        curr_buffer = self._sample_buffers[key]
+        idx = curr_buffer.slot_to_idx[:curr_buffer.count]
+        return idx
+    
+    def get_path_start(self, idx):
+        return self.buffers[self.PATH_START_KEY][idx]
+
+    def get_path_end(self, idx):
+        return self.buffers[self.PATH_END_KEY][idx]
+
+    def get_pathlen(self, idx):
+        is_array = isinstance(idx, np.ndarray) or isinstance(idx, list)
+        if not is_array:
+            idx = [idx]
+
+        n = len(idx)
+        start_idx = self.get_path_start(idx)
+        end_idx = self.get_path_end(idx)
+        pathlen = np.empty(n, dtype=int)
+
+        for i in range(n):
+            curr_start = start_idx[i]
+            curr_end = end_idx[i]
+            if curr_start < curr_end:
+                curr_len = curr_end - curr_start
+            else:
+                curr_len = self.buffer_size - curr_start + curr_end
+            pathlen[i] = curr_len
+
+        if not is_array:
+            pathlen = pathlen[0]
+
+        return pathlen
+
+    def is_valid_path(self, idx):
+        start_idx = self.get_path_start(idx)
+        valid = start_idx != MathUtil.INVALID_IDX
+        return valid
+
+    def store(self, path):
+        start_idx = MathUtil.INVALID_IDX
+        n = path.pathlength()
+        
+        if (n > 0):
+            assert path.is_valid()
+
+            if path.check_vals():
+                if self.buffers is None:
+                    self._init_buffers(path)
+
+                idx = self._request_idx(n + 1)
+                self._store_path(path, idx)
+                self._add_sample_buffers(idx)
+
+                self.num_paths += 1
+                self.total_count += n + 1
+                start_idx = idx[0]
+            else:
+                Logger.print2('Invalid path data value detected')
+        
+        return start_idx
+
+    def clear(self):
+        self.buffer_head = 0
+        self.buffer_tail = MathUtil.INVALID_IDX
+        self.num_paths = 0
+
+        for key in self._sample_buffers:
+            self._sample_buffers[key].clear()
+        return
+
+    def get_next_idx(self, idx):
+        next_idx = np.mod(idx + 1, self.buffer_size)
+        return next_idx
+
+    def is_terminal_state(self, idx):
+        terminate_flags = self.buffers[self.TERMINATE_KEY][idx]
+        terminate = terminate_flags != Env.Terminate.Null.value
+        is_end = self.is_path_end(idx)
+        terminal_state = np.logical_and(terminate, is_end)
+        return terminal_state
+
+    def check_terminal_flag(self, idx, flag):
+        terminate_flags = self.buffers[self.TERMINATE_KEY][idx]
+        terminate = terminate_flags == flag.value
+        return terminate
+
+    def is_path_end(self, idx):
+        is_end = self.buffers[self.PATH_END_KEY][idx] == idx
+        return is_end
+
+    def add_filter_key(self, key):
+        assert self.get_current_size() == 0
+        if key not in self._sample_buffers:
+            self._sample_buffers[key] = SampleBuffer(self.buffer_size)
+        return
+
+    def get_current_size(self):
+        if self.buffer_tail == MathUtil.INVALID_IDX:
+            return 0
+        elif self.buffer_tail < self.buffer_head:
+            return self.buffer_head - self.buffer_tail
+        else:
+            return self.buffer_size - self.buffer_tail + self.buffer_head
+
+    def _check_flags(self, key, flags):
+        return (flags & key) == key
+
+    def _add_sample_buffers(self, idx):
+        flags = self.buffers['flags']
+        for key in self._sample_buffers:
+            curr_buffer = self._sample_buffers[key]
+            filter_idx = [i for i in idx if (self._check_flags(key, flags[i]) and not self.is_path_end(i))]
+            curr_buffer.add(filter_idx)
+        return
+
+    def _free_sample_buffers(self, idx):
+        for key in self._sample_buffers:
+            curr_buffer = self._sample_buffers[key]
+            curr_buffer.free(idx)    
+        return
+
+    def _init_buffers(self, path):
+        self.buffers = dict()
+        self.buffers[self.PATH_START_KEY] = MathUtil.INVALID_IDX * np.ones(self.buffer_size, dtype=int);
+        self.buffers[self.PATH_END_KEY] = MathUtil.INVALID_IDX * np.ones(self.buffer_size, dtype=int);
+
+        for key in dir(path):
+            val = getattr(path, key)
+            if not key.startswith('__') and not inspect.ismethod(val):
+                if key == self.TERMINATE_KEY:
+                    self.buffers[self.TERMINATE_KEY] = np.zeros(shape=[self.buffer_size], dtype=int)
+                else:
+                    val_type = type(val[0])
+                    is_array = val_type == np.ndarray
+                    if is_array:
+                        shape = [self.buffer_size, val[0].shape[0]]
+                        dtype = val[0].dtype
+                    else:
+                        shape = [self.buffer_size]
+                        dtype = val_type
+                    
+                    self.buffers[key] = np.zeros(shape, dtype=dtype)
+        return
+
+    def _request_idx(self, n):
+        assert n + 1 < self.buffer_size # bad things can happen if path is too long
+
+        remainder = n
+        idx = []
+
+        start_idx = self.buffer_head
+        while remainder > 0:
+            end_idx = np.minimum(start_idx + remainder, self.buffer_size)
+            remainder -= (end_idx - start_idx)
+
+            free_idx = list(range(start_idx, end_idx))
+            self._free_idx(free_idx)
+            idx += free_idx
+            start_idx = 0
+
+        self.buffer_head = (self.buffer_head + n) % self.buffer_size
+        return idx
+
+    def _free_idx(self, idx):
+        assert(idx[0] <= idx[-1])
+        n = len(idx)
+        if self.buffer_tail != MathUtil.INVALID_IDX:
+            update_tail = idx[0] <= idx[-1] and idx[0] <= self.buffer_tail and idx[-1] >= self.buffer_tail
+            update_tail |= idx[0] > idx[-1] and (idx[0] <= self.buffer_tail or idx[-1] >= self.buffer_tail)
+            
+            if update_tail:
+                i = 0
+                while i < n:
+                    curr_idx = idx[i]
+                    if self.is_valid_path(curr_idx):
+                        start_idx = self.get_path_start(curr_idx)
+                        end_idx = self.get_path_end(curr_idx)
+                        pathlen = self.get_pathlen(curr_idx)
+
+                        if start_idx < end_idx:
+                            self.buffers[self.PATH_START_KEY][start_idx:end_idx + 1] = MathUtil.INVALID_IDX
+                            self._free_sample_buffers(list(range(start_idx, end_idx + 1)))
+                        else:
+                            self.buffers[self.PATH_START_KEY][start_idx:self.buffer_size] = MathUtil.INVALID_IDX
+                            self.buffers[self.PATH_START_KEY][0:end_idx + 1] = MathUtil.INVALID_IDX
+                            self._free_sample_buffers(list(range(start_idx, self.buffer_size)))
+                            self._free_sample_buffers(list(range(0, end_idx + 1)))
+                        
+                        self.num_paths -= 1
+                        i += pathlen + 1
+                        self.buffer_tail = (end_idx + 1) % self.buffer_size;
+                    else:
+                        i += 1
+        else:
+            self.buffer_tail = idx[0]
+        return
+
+    def _store_path(self, path, idx):
+        n = path.pathlength()
+        for key, data in self.buffers.items():
+            if key != self.PATH_START_KEY and key != self.PATH_END_KEY and key != self.TERMINATE_KEY:
+                val = getattr(path, key)
+                val_len = len(val)
+                assert val_len == n or val_len == n + 1
+                data[idx[:val_len]] = val
+
+        self.buffers[self.TERMINATE_KEY][idx] = path.terminate.value
+        self.buffers[self.PATH_START_KEY][idx] = idx[0]
+        self.buffers[self.PATH_END_KEY][idx] = idx[-1]
+        return
+
+class SampleBuffer(object):
+    def __init__(self, size):
+        self.idx_to_slot = np.empty(shape=[size], dtype=int)
+        self.slot_to_idx = np.empty(shape=[size], dtype=int)
+        self.count = 0
+        self.clear()
+        return
+    
+    def clear(self):
+        self.idx_to_slot.fill(MathUtil.INVALID_IDX)
+        self.slot_to_idx.fill(MathUtil.INVALID_IDX)
+        self.count = 0
+        return
+
+    def is_valid(self, idx):
+        return self.idx_to_slot[idx] != MathUtil.INVALID_IDX
+
+    def get_size(self):
+        return self.idx_to_slot.shape[0]
+
+    def add(self, idx):
+        for i in idx:
+            if not self.is_valid(i):
+                new_slot = self.count
+                assert new_slot >= 0
+
+                self.idx_to_slot[i] = new_slot
+                self.slot_to_idx[new_slot] = i
+                self.count += 1
+        return
+
+    def free(self, idx):
+        for i in idx:
+            if self.is_valid(i):
+                slot = self.idx_to_slot[i]
+                last_slot = self.count - 1
+                last_idx = self.slot_to_idx[last_slot]
+
+                self.idx_to_slot[last_idx] = slot
+                self.slot_to_idx[slot] = last_idx
+                self.idx_to_slot[i] = MathUtil.INVALID_IDX
+                self.slot_to_idx[last_slot] = MathUtil.INVALID_IDX
+                self.count -= 1
+        return
+
+    def sample(self, n):
+        if self.count > 0:
+            slots = np.random.randint(0, self.count, size=n)
+            idx = self.slot_to_idx[slots]
+        else:
+            idx = np.empty(shape=[0], dtype=int)
+        return idx
+
+    def check_consistency(self):
+        valid = True
+        if self.count < 0:
+            valid = False
+
+        if valid:
+            for i in range(self.get_size()):
+                if self.is_valid(i):
+                    s = self.idx_to_slot[i]
+                    if self.slot_to_idx[s] != i:
+                        valid = False
+                        break
+
+                s2i = self.slot_to_idx[i] 
+                if s2i != MathUtil.INVALID_IDX:
+                    i2s = self.idx_to_slot[s2i]
+                    if i2s != i:
+                        valid = False
+                        break
+
+        count0 = np.sum(self.idx_to_slot == MathUtil.INVALID_IDX)
+        count1 = np.sum(self.slot_to_idx == MathUtil.INVALID_IDX)
+        valid &= count0 == count1
+        return valid

examples/pybullet/gym/pybullet_envs/deep_mimic/learning/rl_agent.py

@@ -0,0 +1,595 @@
+import numpy as np
+import copy
+import os
+import time
+
+from abc import ABC, abstractmethod
+from enum import Enum
+
+from learning.path import *
+from learning.exp_params import ExpParams
+from learning.normalizer import Normalizer
+from learning.replay_buffer import ReplayBuffer
+from pybullet_utils.logger import Logger
+import pybullet_utils.mpi_util as MPIUtil
+import pybullet_utils.math_util as MathUtil
+
+class RLAgent(ABC):
+    class Mode(Enum):
+        TRAIN = 0
+        TEST = 1
+        TRAIN_END = 2
+
+    NAME = "None"
+    
+    UPDATE_PERIOD_KEY = "UpdatePeriod"
+    ITERS_PER_UPDATE = "ItersPerUpdate"
+    DISCOUNT_KEY = "Discount"
+    MINI_BATCH_SIZE_KEY = "MiniBatchSize"
+    REPLAY_BUFFER_SIZE_KEY = "ReplayBufferSize"
+    INIT_SAMPLES_KEY = "InitSamples"
+    NORMALIZER_SAMPLES_KEY = "NormalizerSamples"
+
+    OUTPUT_ITERS_KEY = "OutputIters"
+    INT_OUTPUT_ITERS_KEY = "IntOutputIters"
+    TEST_EPISODES_KEY = "TestEpisodes"
+
+    EXP_ANNEAL_SAMPLES_KEY = "ExpAnnealSamples"
+    EXP_PARAM_BEG_KEY = "ExpParamsBeg"
+    EXP_PARAM_END_KEY = "ExpParamsEnd"
+        
+    def __init__(self, world, id, json_data):
+        self.world = world
+        self.id = id
+        self.logger = Logger()
+        self._mode = self.Mode.TRAIN
+        
+        assert self._check_action_space(), \
+            Logger.print2("Invalid action space, got {:s}".format(str(self.get_action_space())))
+        
+        self._enable_training = True
+        self.path = Path()
+        self.iter = int(0)
+        self.start_time = time.time()
+        self._update_counter = 0
+
+        self.update_period = 1.0 # simulated time (seconds) before each training update
+        self.iters_per_update = int(1)
+        self.discount = 0.95
+        self.mini_batch_size = int(32)
+        self.replay_buffer_size = int(50000)
+        self.init_samples = int(1000)
+        self.normalizer_samples = np.inf
+        self._local_mini_batch_size = self.mini_batch_size # batch size for each work for multiprocessing
+        self._need_normalizer_update = True
+        self._total_sample_count = 0
+
+        self._output_dir = ""
+        self._int_output_dir = ""
+        self.output_iters = 100
+        self.int_output_iters = 100
+        
+        self.train_return = 0.0
+        self.test_episodes = int(0)
+        self.test_episode_count = int(0)
+        self.test_return = 0.0
+        self.avg_test_return = 0.0
+        
+        self.exp_anneal_samples = 320000
+        self.exp_params_beg = ExpParams()
+        self.exp_params_end = ExpParams()
+        self.exp_params_curr = ExpParams()
+
+        self._load_params(json_data)
+        self._build_replay_buffer(self.replay_buffer_size)
+        self._build_normalizers()
+        self._build_bounds()
+        self.reset()
+
+        return
+
+    def __str__(self):
+        action_space_str = str(self.get_action_space())
+        info_str = ""
+        info_str += '"ID": {:d},\n "Type": "{:s}",\n "ActionSpace": "{:s}",\n "StateDim": {:d},\n "GoalDim": {:d},\n "ActionDim": {:d}'.format(
+            self.id, self.NAME, action_space_str[action_space_str.rfind('.') + 1:], self.get_state_size(), self.get_goal_size(), self.get_action_size())
+        return "{\n" + info_str + "\n}"
+
+    def get_output_dir(self):
+        return self._output_dir
+    
+    def set_output_dir(self, out_dir):
+        self._output_dir = out_dir
+        if (self._output_dir != ""):
+            self.logger.configure_output_file(out_dir + "/agent" + str(self.id) + "_log.txt")
+        return
+
+    output_dir = property(get_output_dir, set_output_dir)
+
+    def get_int_output_dir(self):
+        return self._int_output_dir
+    
+    def set_int_output_dir(self, out_dir):
+        self._int_output_dir = out_dir
+        return
+
+    int_output_dir = property(get_int_output_dir, set_int_output_dir)
+
+    def reset(self):
+        self.path.clear()
+        return
+
+    def update(self, timestep):
+        if self.need_new_action():
+           #print("update_new_action!!!")
+           self._update_new_action()
+        else:
+           print("no action???")
+
+        if (self._mode == self.Mode.TRAIN and self.enable_training):
+            self._update_counter += timestep
+
+            while self._update_counter >= self.update_period:
+                self._train()
+                self._update_exp_params()
+                self.world.env.set_sample_count(self._total_sample_count)
+                self._update_counter -= self.update_period
+
+        return
+
+    def end_episode(self):
+        if (self.path.pathlength() > 0):
+            self._end_path()
+
+            if (self._mode == self.Mode.TRAIN or self._mode == self.Mode.TRAIN_END):
+                if (self.enable_training and self.path.pathlength() > 0):
+                    self._store_path(self.path)
+            elif (self._mode == self.Mode.TEST):
+                self._update_test_return(self.path)
+            else:
+                assert False, Logger.print2("Unsupported RL agent mode" + str(self._mode))
+
+            self._update_mode()
+        return
+
+    def has_goal(self):
+        return self.get_goal_size() > 0
+
+    def predict_val(self):
+        return 0
+
+    def get_enable_training(self):
+        return self._enable_training
+
+    def set_enable_training(self, enable):
+        print("set_enable_training!=", enable)
+        self._enable_training = enable
+        if (self._enable_training):
+            self.reset()
+        return
+
+    enable_training = property(get_enable_training, set_enable_training)
+
+    def enable_testing(self):
+        return self.test_episodes > 0
+    
+    def get_name(self):
+        return self.NAME
+
+    @abstractmethod
+    def save_model(self, out_path):
+        pass
+
+    @abstractmethod
+    def load_model(self, in_path):
+        pass
+
+    @abstractmethod
+    def _decide_action(self, s, g):
+        pass
+
+    @abstractmethod
+    def _get_output_path(self):
+        pass
+
+    @abstractmethod
+    def _get_int_output_path(self):
+        pass
+
+    @abstractmethod
+    def _train_step(self):
+        pass
+    
+    @abstractmethod
+    def _check_action_space(self):
+        pass
+
+    def get_action_space(self):
+        return self.world.env.get_action_space(self.id)
+
+    def get_state_size(self):
+        return self.world.env.get_state_size(self.id)
+
+    def get_goal_size(self):
+        return self.world.env.get_goal_size(self.id)
+
+    def get_action_size(self):
+        return self.world.env.get_action_size(self.id)
+
+    def get_num_actions(self):
+        return self.world.env.get_num_actions(self.id)
+    
+    def need_new_action(self):
+        return self.world.env.need_new_action(self.id)
+
+    def _build_normalizers(self):
+        self.s_norm = Normalizer(self.get_state_size(), self.world.env.build_state_norm_groups(self.id))
+        self.s_norm.set_mean_std(-self.world.env.build_state_offset(self.id), 
+                                 1 / self.world.env.build_state_scale(self.id))
+
+        self.g_norm = Normalizer(self.get_goal_size(), self.world.env.build_goal_norm_groups(self.id))
+        self.g_norm.set_mean_std(-self.world.env.build_goal_offset(self.id), 
+                                 1 / self.world.env.build_goal_scale(self.id))
+
+        self.a_norm = Normalizer(self.world.env.get_action_size())
+        self.a_norm.set_mean_std(-self.world.env.build_action_offset(self.id), 
+                                 1 / self.world.env.build_action_scale(self.id))
+        return
+
+    def _build_bounds(self):
+        self.a_bound_min = self.world.env.build_action_bound_min(self.id)
+        self.a_bound_max = self.world.env.build_action_bound_max(self.id)
+        return
+
+    def _load_params(self, json_data):
+        if (self.UPDATE_PERIOD_KEY in json_data):
+            self.update_period = int(json_data[self.UPDATE_PERIOD_KEY])
+        
+        if (self.ITERS_PER_UPDATE in json_data):
+            self.iters_per_update = int(json_data[self.ITERS_PER_UPDATE])
+                    
+        if (self.DISCOUNT_KEY in json_data):
+            self.discount = json_data[self.DISCOUNT_KEY]
+        
+        if (self.MINI_BATCH_SIZE_KEY in json_data):
+            self.mini_batch_size = int(json_data[self.MINI_BATCH_SIZE_KEY])
+            
+        if (self.REPLAY_BUFFER_SIZE_KEY in json_data):
+            self.replay_buffer_size = int(json_data[self.REPLAY_BUFFER_SIZE_KEY])
+            
+        if (self.INIT_SAMPLES_KEY in json_data):
+            self.init_samples = int(json_data[self.INIT_SAMPLES_KEY])
+
+        if (self.NORMALIZER_SAMPLES_KEY in json_data):
+            self.normalizer_samples = int(json_data[self.NORMALIZER_SAMPLES_KEY])
+
+        if (self.OUTPUT_ITERS_KEY in json_data):
+            self.output_iters = json_data[self.OUTPUT_ITERS_KEY]
+
+        if (self.INT_OUTPUT_ITERS_KEY in json_data):
+            self.int_output_iters = json_data[self.INT_OUTPUT_ITERS_KEY]
+            
+        if (self.TEST_EPISODES_KEY in json_data):
+            self.test_episodes = int(json_data[self.TEST_EPISODES_KEY])
+
+        if (self.EXP_ANNEAL_SAMPLES_KEY in json_data):
+            self.exp_anneal_samples = json_data[self.EXP_ANNEAL_SAMPLES_KEY]
+
+        if (self.EXP_PARAM_BEG_KEY in json_data):
+            self.exp_params_beg.load(json_data[self.EXP_PARAM_BEG_KEY])
+
+        if (self.EXP_PARAM_END_KEY in json_data):
+            self.exp_params_end.load(json_data[self.EXP_PARAM_END_KEY])
+        
+        num_procs = MPIUtil.get_num_procs()
+        self._local_mini_batch_size = int(np.ceil(self.mini_batch_size / num_procs))
+        self._local_mini_batch_size = np.maximum(self._local_mini_batch_size, 1)
+        self.mini_batch_size = self._local_mini_batch_size * num_procs
+        
+        assert(self.exp_params_beg.noise == self.exp_params_end.noise) # noise std should not change
+        self.exp_params_curr = copy.deepcopy(self.exp_params_beg)
+        self.exp_params_end.noise = self.exp_params_beg.noise
+
+        self._need_normalizer_update = self.normalizer_samples > 0
+
+        return
+
+    def _record_state(self):
+        s = self.world.env.record_state(self.id)
+        return s
+
+    def _record_goal(self):
+        g = self.world.env.record_goal(self.id)
+        return g
+
+    def _record_reward(self):
+        r = self.world.env.calc_reward(self.id)
+        return r
+
+    def _apply_action(self, a):
+        self.world.env.set_action(self.id, a)
+        return
+
+    def _record_flags(self):
+        return int(0)
+
+    def _is_first_step(self):
+        return len(self.path.states) == 0
+
+    def _end_path(self):
+        s = self._record_state()
+        g = self._record_goal()
+        r = self._record_reward()
+
+        self.path.rewards.append(r)
+        self.path.states.append(s)
+        self.path.goals.append(g)
+        self.path.terminate = self.world.env.check_terminate(self.id)
+
+        return
+
+    def _update_new_action(self):
+        s = self._record_state()
+        g = self._record_goal()
+
+        if not (self._is_first_step()):
+            r = self._record_reward()
+            self.path.rewards.append(r)
+        
+        a, logp = self._decide_action(s=s, g=g)
+        assert len(np.shape(a)) == 1
+        assert len(np.shape(logp)) <= 1
+
+        flags = self._record_flags()
+        self._apply_action(a)
+
+        self.path.states.append(s)
+        self.path.goals.append(g)
+        self.path.actions.append(a)
+        self.path.logps.append(logp)
+        self.path.flags.append(flags)
+        
+        if self._enable_draw():
+            self._log_val(s, g)
+        
+        return
+    
+    def _update_exp_params(self):
+        lerp = float(self._total_sample_count) / self.exp_anneal_samples
+        lerp = np.clip(lerp, 0.0, 1.0)
+        self.exp_params_curr = self.exp_params_beg.lerp(self.exp_params_end, lerp)
+        return
+
+    def _update_test_return(self, path):
+        path_reward = path.calc_return()
+        self.test_return += path_reward
+        self.test_episode_count += 1
+        return
+
+    def _update_mode(self):
+        if (self._mode == self.Mode.TRAIN):
+            self._update_mode_train()
+        elif (self._mode == self.Mode.TRAIN_END):
+            self._update_mode_train_end()
+        elif (self._mode == self.Mode.TEST):
+            self._update_mode_test()
+        else:
+            assert False, Logger.print2("Unsupported RL agent mode" + str(self._mode))
+        return
+
+    def _update_mode_train(self):
+        return
+
+    def _update_mode_train_end(self):
+        self._init_mode_test()
+        return
+
+    def _update_mode_test(self):
+        if (self.test_episode_count * MPIUtil.get_num_procs() >= self.test_episodes):
+            global_return = MPIUtil.reduce_sum(self.test_return)
+            global_count = MPIUtil.reduce_sum(self.test_episode_count)
+            avg_return = global_return / global_count
+            self.avg_test_return = avg_return
+
+            if self.enable_training:
+                self._init_mode_train()
+        return
+
+    def _init_mode_train(self):
+        self._mode = self.Mode.TRAIN
+        self.world.env.set_mode(self._mode)
+        return
+
+    def _init_mode_train_end(self):
+        self._mode = self.Mode.TRAIN_END
+        return
+
+    def _init_mode_test(self):
+        self._mode = self.Mode.TEST
+        self.test_return = 0.0
+        self.test_episode_count = 0
+        self.world.env.set_mode(self._mode)
+        return
+
+    def _enable_output(self):
+        return MPIUtil.is_root_proc() and self.output_dir != ""
+
+    def _enable_int_output(self):
+        return MPIUtil.is_root_proc() and self.int_output_dir != ""
+    
+    def _calc_val_bounds(self, discount):
+        r_min = self.world.env.get_reward_min(self.id)
+        r_max = self.world.env.get_reward_max(self.id)
+        assert(r_min <= r_max)
+
+        val_min = r_min / ( 1.0 - discount)
+        val_max = r_max / ( 1.0 - discount)
+        return val_min, val_max
+
+    def _calc_val_offset_scale(self, discount):
+        val_min, val_max = self._calc_val_bounds(discount)
+        val_offset = 0
+        val_scale = 1
+
+        if (np.isfinite(val_min) and np.isfinite(val_max)):
+            val_offset = -0.5 * (val_max + val_min)
+            val_scale = 2 / (val_max - val_min)
+
+        return val_offset, val_scale
+
+    def _calc_term_vals(self, discount):
+        r_fail = self.world.env.get_reward_fail(self.id)
+        r_succ = self.world.env.get_reward_succ(self.id)
+
+        r_min = self.world.env.get_reward_min(self.id)
+        r_max = self.world.env.get_reward_max(self.id)
+        assert(r_fail <= r_max and r_fail >= r_min)
+        assert(r_succ <= r_max and r_succ >= r_min)
+        assert(not np.isinf(r_fail))
+        assert(not np.isinf(r_succ))
+
+        if (discount == 0):
+            val_fail = 0
+            val_succ = 0
+        else:
+            val_fail = r_fail / (1.0 - discount)
+            val_succ = r_succ / (1.0 - discount)
+
+        return val_fail, val_succ
+    
+    def _update_iter(self, iter):
+        if (self._enable_output() and self.iter % self.output_iters == 0):
+            output_path = self._get_output_path()
+            output_dir = os.path.dirname(output_path)
+            if not os.path.exists(output_dir):
+                os.makedirs(output_dir)
+            self.save_model(output_path)
+
+        if (self._enable_int_output() and self.iter % self.int_output_iters == 0):
+            int_output_path = self._get_int_output_path()
+            int_output_dir = os.path.dirname(int_output_path)
+            if not os.path.exists(int_output_dir):
+                os.makedirs(int_output_dir)
+            self.save_model(int_output_path)
+
+        self.iter = iter
+        return
+
+    def _enable_draw(self):
+        return self.world.env.enable_draw
+
+    def _log_val(self, s, g):
+        pass
+
+    def _build_replay_buffer(self, buffer_size):
+        num_procs = MPIUtil.get_num_procs()
+        buffer_size = int(buffer_size / num_procs)
+        self.replay_buffer = ReplayBuffer(buffer_size=buffer_size)
+        self.replay_buffer_initialized = False
+        return
+    
+    def _store_path(self, path):
+        path_id = self.replay_buffer.store(path)
+        valid_path = path_id != MathUtil.INVALID_IDX
+
+        if valid_path:
+            self.train_return = path.calc_return()
+
+            if self._need_normalizer_update:
+                self._record_normalizers(path)
+
+        return path_id
+
+    def _record_normalizers(self, path):
+        states = np.array(path.states)
+        self.s_norm.record(states)
+
+        if self.has_goal():
+            goals = np.array(path.goals)
+            self.g_norm.record(goals)
+
+        return
+
+    def _update_normalizers(self):
+        self.s_norm.update()
+
+        if self.has_goal():
+            self.g_norm.update()
+        return
+
+    def _train(self):
+        samples = self.replay_buffer.total_count
+        self._total_sample_count = int(MPIUtil.reduce_sum(samples))
+        end_training = False
+        
+        if (self.replay_buffer_initialized):  
+            if (self._valid_train_step()):
+                prev_iter = self.iter
+                iters = self._get_iters_per_update()
+                avg_train_return = MPIUtil.reduce_avg(self.train_return)
+            
+                for i in range(iters):
+                    curr_iter = self.iter
+                    wall_time = time.time() - self.start_time
+                    wall_time /= 60 * 60 # store time in hours
+
+                    has_goal = self.has_goal()
+                    s_mean = np.mean(self.s_norm.mean)
+                    s_std = np.mean(self.s_norm.std)
+                    g_mean = np.mean(self.g_norm.mean) if has_goal else 0
+                    g_std = np.mean(self.g_norm.std) if has_goal else 0
+
+                    self.logger.log_tabular("Iteration", self.iter)
+                    self.logger.log_tabular("Wall_Time", wall_time)
+                    self.logger.log_tabular("Samples", self._total_sample_count)
+                    self.logger.log_tabular("Train_Return", avg_train_return)
+                    self.logger.log_tabular("Test_Return", self.avg_test_return)
+                    self.logger.log_tabular("State_Mean", s_mean)
+                    self.logger.log_tabular("State_Std", s_std)
+                    self.logger.log_tabular("Goal_Mean", g_mean)
+                    self.logger.log_tabular("Goal_Std", g_std)
+                    self._log_exp_params()
+
+                    self._update_iter(self.iter + 1)
+                    self._train_step()
+
+                    Logger.print2("Agent " + str(self.id))
+                    self.logger.print_tabular()
+                    Logger.print2("") 
+
+                    if (self._enable_output() and curr_iter % self.int_output_iters == 0):
+                        self.logger.dump_tabular()
+
+                if (prev_iter // self.int_output_iters != self.iter // self.int_output_iters):
+                    end_training = self.enable_testing()
+
+        else:
+
+            Logger.print2("Agent " + str(self.id))
+            Logger.print2("Samples: " + str(self._total_sample_count))
+            Logger.print2("") 
+
+            if (self._total_sample_count >= self.init_samples):
+                self.replay_buffer_initialized = True
+                end_training = self.enable_testing()
+        
+        if self._need_normalizer_update:
+            self._update_normalizers()
+            self._need_normalizer_update = self.normalizer_samples > self._total_sample_count
+
+        if end_training:
+            self._init_mode_train_end()
+ 
+        return
+
+    def _get_iters_per_update(self):
+        return MPIUtil.get_num_procs() * self.iters_per_update
+
+    def _valid_train_step(self):
+        return True 
+
+    def _log_exp_params(self):
+        self.logger.log_tabular("Exp_Rate", self.exp_params_curr.rate)
+        self.logger.log_tabular("Exp_Noise", self.exp_params_curr.noise)
+        self.logger.log_tabular("Exp_Temp", self.exp_params_curr.temp)
+        return

examples/pybullet/gym/pybullet_envs/deep_mimic/learning/rl_util.py

@@ -0,0 +1,18 @@
+import numpy as np
+
+def compute_return(rewards, gamma, td_lambda, val_t):
+    # computes td-lambda return of path
+    path_len = len(rewards)
+    assert len(val_t) == path_len + 1
+
+    return_t = np.zeros(path_len)
+    last_val = rewards[-1] + gamma * val_t[-1]
+    return_t[-1] = last_val
+
+    for i in reversed(range(0, path_len - 1)):
+        curr_r = rewards[i]
+        next_ret = return_t[i + 1]
+        curr_val = curr_r + gamma * ((1.0 - td_lambda) * val_t[i + 1] + td_lambda * next_ret)
+        return_t[i] = curr_val
+    
+    return return_t

examples/pybullet/gym/pybullet_envs/deep_mimic/learning/rl_world.py

@@ -0,0 +1,142 @@
+import numpy as np
+import learning.agent_builder as AgentBuilder
+import learning.tf_util as TFUtil
+from learning.rl_agent import RLAgent
+from pybullet_utils.logger import Logger
+
+class RLWorld(object):
+    def __init__(self, env, arg_parser):
+        TFUtil.disable_gpu()
+
+        self.env = env
+        self.arg_parser = arg_parser
+        self._enable_training = True
+        self.train_agents = []
+        self.parse_args(arg_parser)
+
+        self.build_agents()
+        
+        return
+
+    def get_enable_training(self):
+        return self._enable_training
+    
+    def set_enable_training(self, enable):
+        self._enable_training = enable
+        for i in range(len(self.agents)):
+            curr_agent = self.agents[i]
+            if curr_agent is not None:
+                enable_curr_train = self.train_agents[i] if (len(self.train_agents) > 0) else True
+                curr_agent.enable_training = self.enable_training and enable_curr_train
+
+        if (self._enable_training):
+            self.env.set_mode(RLAgent.Mode.TRAIN)
+        else:
+            self.env.set_mode(RLAgent.Mode.TEST)
+
+        return
+
+    enable_training = property(get_enable_training, set_enable_training)
+    
+    def parse_args(self, arg_parser):
+        self.train_agents = self.arg_parser.parse_bools('train_agents')
+        num_agents = self.env.get_num_agents()
+        assert(len(self.train_agents) == num_agents or len(self.train_agents) == 0)
+
+        return
+
+    def shutdown(self):
+        self.env.shutdown()
+        return
+
+    def build_agents(self):
+        num_agents = self.env.get_num_agents()
+        print("num_agents=",num_agents)
+        self.agents = []
+
+        Logger.print2('')
+        Logger.print2('Num Agents: {:d}'.format(num_agents))
+
+        agent_files = self.arg_parser.parse_strings('agent_files')
+        print("len(agent_files)=",len(agent_files))
+        assert(len(agent_files) == num_agents or len(agent_files) == 0)
+
+        model_files = self.arg_parser.parse_strings('model_files')
+        assert(len(model_files) == num_agents or len(model_files) == 0)
+
+        output_path = self.arg_parser.parse_string('output_path')
+        int_output_path = self.arg_parser.parse_string('int_output_path')
+
+        for i in range(num_agents):
+            curr_file = agent_files[i]
+            curr_agent = self._build_agent(i, curr_file)
+
+            if curr_agent is not None:
+                curr_agent.output_dir = output_path
+                curr_agent.int_output_dir = int_output_path
+                Logger.print2(str(curr_agent))
+
+                if (len(model_files) > 0):
+                    curr_model_file = model_files[i]
+                    if curr_model_file != 'none':
+                        curr_agent.load_model(curr_model_file)
+
+            self.agents.append(curr_agent)
+            Logger.print2('')
+
+        self.set_enable_training(self.enable_training)
+
+        return
+
+    def update(self, timestep):
+        #print("world update!\n")
+        self._update_agents(timestep)
+        self._update_env(timestep)
+        return
+
+    def reset(self):
+        self._reset_agents()
+        self._reset_env()
+        return
+
+    def end_episode(self):
+        self._end_episode_agents();
+        return
+
+    def _update_env(self, timestep):
+        self.env.update(timestep)
+        return
+
+    def _update_agents(self, timestep):
+        #print("len(agents)=",len(self.agents))
+        for agent in self.agents:
+            if (agent is not None):
+                agent.update(timestep)
+        return
+
+    def _reset_env(self):
+        self.env.reset()
+        return
+
+    def _reset_agents(self):
+        for agent in self.agents:
+            if (agent != None):
+                agent.reset()
+        return
+
+    def _end_episode_agents(self):
+        for agent in self.agents:
+            if (agent != None):
+                agent.end_episode()
+        return
+
+    def _build_agent(self, id, agent_file):
+        Logger.print2('Agent {:d}: {}'.format(id, agent_file))
+        if (agent_file == 'none'):
+            agent = None
+        else:
+            agent = AgentBuilder.build_agent(self, id, agent_file)
+            assert (agent != None), 'Failed to build agent {:d} from: {}'.format(id, agent_file)
+        
+        return agent
+        

examples/pybullet/gym/pybullet_envs/deep_mimic/learning/solvers/mpi_solver.py

@@ -0,0 +1,103 @@
+from mpi4py import MPI
+import tensorflow as tf
+import numpy as np
+import learning.tf_util as TFUtil
+import pybullet_utils.math_util as MathUtil
+import pybullet_utils.mpi_util as MPIUtil
+from pybullet_utils.logger import Logger
+
+from learning.solvers.solver import Solver
+
+class MPISolver(Solver):
+    CHECK_SYNC_ITERS = 1000
+
+    def __init__(self, sess, optimizer, vars):
+        super().__init__(vars)
+        self.sess = sess
+        self.optimizer = optimizer
+        self._build_grad_feed(vars)
+        self._update = optimizer.apply_gradients(zip(self._grad_tf_list, self.vars))
+        self._set_flat_vars = TFUtil.SetFromFlat(sess, self.vars)
+        self._get_flat_vars = TFUtil.GetFlat(sess, self.vars)
+
+        self.iter = 0
+        grad_dim = self._calc_grad_dim()
+        self._flat_grad = np.zeros(grad_dim, dtype=np.float32)
+        self._global_flat_grad = np.zeros(grad_dim, dtype=np.float32)
+        
+        return
+
+    def get_stepsize(self):
+        return self.optimizer._learning_rate_tensor.eval()
+
+    def update(self, grads=None, grad_scale=1.0):
+        if grads is not None:
+            self._flat_grad = MathUtil.flatten(grads)
+        else:
+            self._flat_grad.fill(0)
+        return self.update_flatgrad(self._flat_grad, grad_scale)
+
+    def update_flatgrad(self, flat_grad, grad_scale=1.0):
+        if self.iter % self.CHECK_SYNC_ITERS == 0:
+            assert self.check_synced(), Logger.print2('Network parameters desynchronized')
+        
+        if grad_scale != 1.0:
+            flat_grad *= grad_scale
+
+        MPI.COMM_WORLD.Allreduce(flat_grad, self._global_flat_grad, op=MPI.SUM)
+        self._global_flat_grad /= MPIUtil.get_num_procs()
+
+        self._load_flat_grad(self._global_flat_grad)
+        self.sess.run([self._update], self._grad_feed)
+        self.iter += 1
+
+        return
+
+    def sync(self):
+        vars = self._get_flat_vars()
+        MPIUtil.bcast(vars)
+        self._set_flat_vars(vars)
+        return
+
+    def check_synced(self):
+        synced = True
+        if self._is_root():
+            vars = self._get_flat_vars()
+            MPIUtil.bcast(vars)
+        else:
+            vars_local = self._get_flat_vars()
+            vars_root = np.empty_like(vars_local)
+            MPIUtil.bcast(vars_root)
+            synced = (vars_local == vars_root).all()
+        return synced
+
+    def _is_root(self):
+        return MPIUtil.is_root_proc()
+    
+    def _build_grad_feed(self, vars):
+        self._grad_tf_list = []
+        self._grad_buffers = []
+        for v in self.vars:
+            shape = v.get_shape()
+            grad = np.zeros(shape)
+            grad_tf = tf.placeholder(tf.float32, shape=shape)
+            self._grad_buffers.append(grad)
+            self._grad_tf_list.append(grad_tf)
+
+        self._grad_feed = dict({g_tf: g for g_tf, g in zip(self._grad_tf_list, self._grad_buffers)})
+        
+        return
+
+    def _calc_grad_dim(self):
+        grad_dim = 0
+        for grad in self._grad_buffers:
+            grad_dim += grad.size
+        return grad_dim
+
+    def _load_flat_grad(self, flat_grad):
+        start = 0
+        for g in self._grad_buffers:
+            size = g.size
+            np.copyto(g, np.reshape(flat_grad[start:start + size], g.shape))
+            start += size
+        return

examples/pybullet/gym/pybullet_envs/deep_mimic/learning/solvers/solver.py

@@ -0,0 +1,10 @@
+from abc import ABC, abstractmethod
+
+class Solver(ABC):
+    def __init__(self, vars):
+        self.vars = vars
+        return
+
+    @abstractmethod
+    def update(self, grads):
+        pass

examples/pybullet/gym/pybullet_envs/deep_mimic/learning/tf_agent.py

@@ -0,0 +1,149 @@
+import numpy as np
+import tensorflow as tf
+from abc import abstractmethod
+
+from learning.rl_agent import RLAgent
+from pybullet_utils.logger import Logger
+from learning.tf_normalizer import TFNormalizer
+
+class TFAgent(RLAgent):
+    RESOURCE_SCOPE = 'resource'
+    SOLVER_SCOPE = 'solvers'
+
+    def __init__(self, world, id, json_data):
+        self.tf_scope = 'agent'
+        self.graph = tf.Graph()
+        self.sess = tf.Session(graph=self.graph)
+
+        super().__init__(world, id, json_data)
+        self._build_graph(json_data)
+        self._init_normalizers()
+        return
+
+    def __del__(self):
+        self.sess.close()
+        return
+
+    def save_model(self, out_path):
+        with self.sess.as_default(), self.graph.as_default():
+            try:
+                save_path = self.saver.save(self.sess, out_path, write_meta_graph=False, write_state=False)
+                Logger.print2('Model saved to: ' + save_path)
+            except:
+                Logger.print2("Failed to save model to: " + save_path)
+        return
+
+    def load_model(self, in_path):
+        with self.sess.as_default(), self.graph.as_default():
+            self.saver.restore(self.sess, in_path)
+            self._load_normalizers()
+            Logger.print2('Model loaded from: ' + in_path)
+        return
+
+    def _get_output_path(self):
+        assert(self.output_dir != '')
+        file_path = self.output_dir + '/agent' + str(self.id) + '_model.ckpt'
+        return file_path
+
+    def _get_int_output_path(self):
+        assert(self.int_output_dir != '')
+        file_path = self.int_output_dir + ('/agent{:d}_models/agent{:d}_int_model_{:010d}.ckpt').format(self.id, self.id, self.iter)
+        return file_path
+
+    def _build_graph(self, json_data):
+        with self.sess.as_default(), self.graph.as_default():
+            with tf.variable_scope(self.tf_scope):
+                self._build_nets(json_data)
+                
+                with tf.variable_scope(self.SOLVER_SCOPE):
+                    self._build_losses(json_data)
+                    self._build_solvers(json_data)
+
+                self._initialize_vars()
+                self._build_saver()
+        return
+
+    def _init_normalizers(self):
+        with self.sess.as_default(), self.graph.as_default():
+            # update normalizers to sync the tensorflow tensors
+            self.s_norm.update()
+            self.g_norm.update()
+            self.a_norm.update()
+        return
+
+    @abstractmethod
+    def _build_nets(self, json_data):
+        pass
+
+    @abstractmethod
+    def _build_losses(self, json_data):
+        pass
+
+    @abstractmethod
+    def _build_solvers(self, json_data):
+        pass
+
+    def _tf_vars(self, scope=''):
+        with self.sess.as_default(), self.graph.as_default():
+            res = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=self.tf_scope + '/' + scope)
+            assert len(res) > 0
+        return res
+
+    def _build_normalizers(self):
+        with self.sess.as_default(), self.graph.as_default(), tf.variable_scope(self.tf_scope):
+            with tf.variable_scope(self.RESOURCE_SCOPE):
+                self.s_norm = TFNormalizer(self.sess, 's_norm', self.get_state_size(), self.world.env.build_state_norm_groups(self.id))
+                state_offset = -self.world.env.build_state_offset(self.id)
+                print("state_offset=",state_offset)
+                state_scale = 1 / self.world.env.build_state_scale(self.id)
+                print("state_scale=",state_scale)
+                self.s_norm.set_mean_std(-self.world.env.build_state_offset(self.id), 
+                                         1 / self.world.env.build_state_scale(self.id))
+                
+                self.g_norm = TFNormalizer(self.sess, 'g_norm', self.get_goal_size(), self.world.env.build_goal_norm_groups(self.id))
+                self.g_norm.set_mean_std(-self.world.env.build_goal_offset(self.id), 
+                                         1 / self.world.env.build_goal_scale(self.id))
+
+                self.a_norm = TFNormalizer(self.sess, 'a_norm', self.get_action_size())
+                self.a_norm.set_mean_std(-self.world.env.build_action_offset(self.id), 
+                                         1 / self.world.env.build_action_scale(self.id))
+        return
+
+    def _load_normalizers(self):
+        self.s_norm.load()
+        self.g_norm.load()
+        self.a_norm.load()
+        return
+
+    def _update_normalizers(self):
+        with self.sess.as_default(), self.graph.as_default():
+            super()._update_normalizers()
+        return
+
+    def _initialize_vars(self):
+        self.sess.run(tf.global_variables_initializer())
+        return
+
+    def _build_saver(self):
+        vars = self._get_saver_vars()
+        self.saver = tf.train.Saver(vars, max_to_keep=0)
+        return
+
+    def _get_saver_vars(self):
+        with self.sess.as_default(), self.graph.as_default():
+            vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=self.tf_scope)
+            vars = [v for v in vars if '/' + self.SOLVER_SCOPE + '/' not in v.name]
+            #vars = [v for v in vars if '/target/' not in v.name]
+            assert len(vars) > 0
+        return vars
+    
+    def _weight_decay_loss(self, scope):
+        vars = self._tf_vars(scope)
+        vars_no_bias = [v for v in vars if 'bias' not in v.name]
+        loss = tf.add_n([tf.nn.l2_loss(v) for v in vars_no_bias])
+        return loss
+
+    def _train(self):
+        with self.sess.as_default(), self.graph.as_default():
+            super()._train()
+        return

examples/pybullet/gym/pybullet_envs/deep_mimic/learning/tf_normalizer.py

@@ -0,0 +1,67 @@
+import numpy as np
+import copy
+import tensorflow as tf
+from learning.normalizer import Normalizer
+
+class TFNormalizer(Normalizer):
+
+    def __init__(self, sess, scope, size, groups_ids=None, eps=0.02, clip=np.inf):
+        self.sess = sess
+        self.scope = scope
+        super().__init__(size, groups_ids, eps, clip)
+
+        with tf.variable_scope(self.scope):
+            self._build_resource_tf()
+        return
+
+    # initialze count when loading saved values so that things don't change to quickly during updates
+    def load(self):
+        self.count = self.count_tf.eval()[0]
+        self.mean = self.mean_tf.eval()
+        self.std = self.std_tf.eval()
+        self.mean_sq = self.calc_mean_sq(self.mean, self.std)
+        return
+
+    def update(self):
+        super().update()
+        self._update_resource_tf()
+        return
+
+    def set_mean_std(self, mean, std):
+        super().set_mean_std(mean, std)
+        self._update_resource_tf()
+        return
+
+    def normalize_tf(self, x):
+        norm_x = (x - self.mean_tf) / self.std_tf
+        norm_x = tf.clip_by_value(norm_x, -self.clip, self.clip)
+        return norm_x
+
+    def unnormalize_tf(self, norm_x):
+        x = norm_x * self.std_tf + self.mean_tf
+        return x
+    
+    def _build_resource_tf(self):
+        self.count_tf = tf.get_variable(dtype=tf.int32, name='count', initializer=np.array([self.count], dtype=np.int32), trainable=False)
+        self.mean_tf = tf.get_variable(dtype=tf.float32, name='mean', initializer=self.mean.astype(np.float32), trainable=False)
+        self.std_tf = tf.get_variable(dtype=tf.float32, name='std', initializer=self.std.astype(np.float32), trainable=False)
+        
+        self.count_ph = tf.get_variable(dtype=tf.int32, name='count_ph', shape=[1])
+        self.mean_ph = tf.get_variable(dtype=tf.float32, name='mean_ph', shape=self.mean.shape)
+        self.std_ph = tf.get_variable(dtype=tf.float32, name='std_ph', shape=self.std.shape)
+        
+        self._update_op = tf.group(
+            self.count_tf.assign(self.count_ph),
+            self.mean_tf.assign(self.mean_ph),
+            self.std_tf.assign(self.std_ph)
+        )
+        return
+
+    def _update_resource_tf(self):
+        feed = {
+            self.count_ph: np.array([self.count], dtype=np.int32),
+            self.mean_ph: self.mean,
+            self.std_ph: self.std
+        }
+        self.sess.run(self._update_op, feed_dict=feed)
+        return

examples/pybullet/gym/pybullet_envs/deep_mimic/learning/tf_util.py

@@ -0,0 +1,104 @@
+import tensorflow as tf
+import numpy as np
+import os
+
+xavier_initializer = tf.contrib.layers.xavier_initializer()
+
+def disable_gpu():
+    os.environ["CUDA_VISIBLE_DEVICES"] = '-1'
+    return
+
+def var_shape(x):
+    out = [k.value for k in x.get_shape()]
+    assert all(isinstance(a, int) for a in out), "shape function assumes that shape is fully known"
+    return out
+
+def intprod(x):
+    return int(np.prod(x))
+
+def numel(x):
+    n = intprod(var_shape(x))
+    return n
+
+def flat_grad(loss, var_list, grad_ys=None):
+    grads = tf.gradients(loss, var_list, grad_ys)
+    return tf.concat([tf.reshape(grad, [numel(v)]) for (v, grad) in zip(var_list, grads)], axis=0)
+
+def fc_net(input, layers_sizes, activation, reuse=None, flatten=False): # build fully connected network
+    curr_tf = input
+    for i, size in enumerate(layers_sizes):
+        with tf.variable_scope(str(i), reuse=reuse):
+            curr_tf = tf.layers.dense(inputs=curr_tf,
+                                    units=size,
+                                    kernel_initializer=xavier_initializer,
+                                    activation = activation if i < len(layers_sizes)-1 else None)
+    if flatten:
+        assert layers_sizes[-1] == 1
+        curr_tf = tf.reshape(curr_tf, [-1])
+
+    return curr_tf
+
+def copy(sess, src, dst):
+    assert len(src) == len(dst)
+    sess.run(list(map(lambda v: v[1].assign(v[0]), zip(src, dst))))
+    return
+
+def flat_grad(loss, var_list):
+    grads = tf.gradients(loss, var_list)
+    return tf.concat(axis=0, values=[tf.reshape(grad, [numel(v)])
+        for (v, grad) in zip(var_list, grads)])
+
+
+def calc_logp_gaussian(x_tf, mean_tf, std_tf):
+    dim = tf.to_float(tf.shape(x_tf)[-1])
+
+    if mean_tf is None:
+        diff_tf = x_tf
+    else:
+        diff_tf = x_tf - mean_tf
+
+    logp_tf = -0.5 * tf.reduce_sum(tf.square(diff_tf / std_tf), axis=-1)
+    logp_tf += -0.5 * dim * np.log(2 * np.pi) - tf.reduce_sum(tf.log(std_tf), axis=-1)
+    
+    return logp_tf
+
+def calc_bound_loss(x_tf, bound_min, bound_max):
+    # penalty for violating bounds
+    violation_min = tf.minimum(x_tf - bound_min, 0)
+    violation_max = tf.maximum(x_tf - bound_max, 0)
+    violation = tf.reduce_sum(tf.square(violation_min), axis=-1) + tf.reduce_sum(tf.square(violation_max), axis=-1)
+    loss = 0.5 * tf.reduce_mean(violation)
+    return loss
+
+class SetFromFlat(object):
+    def __init__(self, sess, var_list, dtype=tf.float32):
+        assigns = []
+        shapes = list(map(var_shape, var_list))
+        total_size = np.sum([intprod(shape) for shape in shapes])
+
+        self.sess = sess
+        self.theta = tf.placeholder(dtype,[total_size])
+        start=0
+        assigns = []
+
+        for (shape,v) in zip(shapes,var_list):
+            size = intprod(shape)
+            assigns.append(tf.assign(v, tf.reshape(self.theta[start:start+size],shape)))
+            start += size
+
+        self.op = tf.group(*assigns)
+
+        return
+
+    def __call__(self, theta):
+        self.sess.run(self.op, feed_dict={self.theta:theta})
+        return
+
+class GetFlat(object):
+    def __init__(self, sess, var_list):
+        self.sess = sess
+        self.op = tf.concat(axis=0, values=[tf.reshape(v, [numel(v)]) for v in var_list])
+        return
+
+    def __call__(self):
+        return self.sess.run(self.op)

examples/pybullet/gym/pybullet_envs/deep_mimic/testrl.py

@@ -0,0 +1,65 @@
+import json
+from learning.rl_world import RLWorld
+from learning.ppo_agent import PPOAgent
+
+import pybullet_data
+from pybullet_utils.arg_parser import ArgParser
+from pybullet_utils.logger import Logger
+from pybullet_envs.deep_mimic.env.pybullet_deep_mimic_env import PyBulletDeepMimicEnv
+import sys
+import random
+
+def build_arg_parser(args):
+    arg_parser = ArgParser()
+    arg_parser.load_args(args)
+
+    arg_file = arg_parser.parse_string('arg_file', '')
+    if (arg_file != ''):
+        path = pybullet_data.getDataPath()+"/args/"+arg_file
+        succ = arg_parser.load_file(path)
+        Logger.print2(arg_file)
+        assert succ, Logger.print2('Failed to load args from: ' + arg_file)
+    return arg_parser
+
+args = sys.argv[1:]
+arg_parser = build_arg_parser(args)
+
+render=True
+env = PyBulletDeepMimicEnv (args,render)
+ 
+world = RLWorld(env, arg_parser)
+
+motion_file = arg_parser.parse_string("motion_file")
+print("motion_file=",motion_file)
+bodies = arg_parser.parse_ints("fall_contact_bodies")
+print("bodies=",bodies)
+int_output_path = arg_parser.parse_string("int_output_path")
+print("int_output_path=",int_output_path)
+
+agent_files = arg_parser.parse_string("agent_files")
+
+AGENT_TYPE_KEY = "AgentType"
+
+print("agent_file=",agent_files)
+with open(agent_files) as data_file:
+    json_data = json.load(data_file)
+    print("json_data=",json_data)
+    assert AGENT_TYPE_KEY in json_data
+    agent_type = json_data[AGENT_TYPE_KEY]
+    print("agent_type=",agent_type)
+    agent = PPOAgent(world, id, json_data)
+    
+    agent.set_enable_training(True)
+    world.reset()
+
+    while (world.env._pybullet_client.isConnected()):
+
+      timeStep = 1./600.
+      world.update(timeStep)
+      reward = world.env.calc_reward(agent_id=0)
+      #print("reward=",reward)
+      
+      end_episode = world.env.is_episode_end()
+      if (end_episode):
+        world.end_episode()
+        world.reset()

examples/pybullet/premake4.lua

@@ -17,7 +17,9 @@ project ("pybullet")
 		
 		includedirs {"../../src", "../../examples",
 		"../../examples/ThirdPartyLibs"}
-		defines {"PHYSICS_IN_PROCESS_EXAMPLE_BROWSER"}
+		defines {"PHYSICS_IN_PROCESS_EXAMPLE_BROWSER", "STATIC_LINK_SPD_PLUGIN"}
+		
+		
 	hasCL = findOpenCL("clew")
 
 	links{ "BulletExampleBrowserLib","gwen", "BulletFileLoader","BulletWorldImporter","OpenGL_Window","BulletSoftBody", "BulletInverseDynamicsUtils", "BulletInverseDynamics", "BulletDynamics","BulletCollision","LinearMath","BussIK", "Bullet3Common"}
@@ -170,6 +172,20 @@ if not _OPTIONS["no-enet"] then
 			"../../examples/SharedMemory/plugins/collisionFilterPlugin/collisionFilterPlugin.cpp",
 			"../../examples/SharedMemory/plugins/pdControlPlugin/pdControlPlugin.cpp",
 			"../../examples/SharedMemory/plugins/pdControlPlugin/pdControlPlugin.h",
+			"../../examples/SharedMemory/plugins/stablePDPlugin/SpAlg.cpp",
+			"../../examples/SharedMemory/plugins/stablePDPlugin/SpAlg.h",
+			"../../examples/SharedMemory/plugins/stablePDPlugin/Shape.cpp",
+			"../../examples/SharedMemory/plugins/stablePDPlugin/Shape.h",
+			"../../examples/SharedMemory/plugins/stablePDPlugin/RBDUtil.cpp",
+			"../../examples/SharedMemory/plugins/stablePDPlugin/RBDUtil.h",
+			"../../examples/SharedMemory/plugins/stablePDPlugin/RBDModel.cpp",
+			"../../examples/SharedMemory/plugins/stablePDPlugin/RBDModel.h",
+			"../../examples/SharedMemory/plugins/stablePDPlugin/MathUtil.cpp",
+			"../../examples/SharedMemory/plugins/stablePDPlugin/MathUtil.h",
+			"../../examples/SharedMemory/plugins/stablePDPlugin/KinTree.cpp",
+			"../../examples/SharedMemory/plugins/stablePDPlugin/KinTree.h",
+			"../../examples/SharedMemory/plugins/stablePDPlugin/BulletConversion.cpp",
+			"../../examples/SharedMemory/plugins/stablePDPlugin/BulletConversion.h",
 			}
 
 	if _OPTIONS["enable_physx"] then

src/Bullet3Common/b3Quaternion.h

@@ -92,8 +92,11 @@ public:
 	/**@brief Set the rotation using axis angle notation 
    * @param axis The axis around which to rotate
    * @param angle The magnitude of the rotation in Radians */
-	void setRotation(const b3Vector3& axis, const b3Scalar& _angle)
+	void setRotation(const b3Vector3& axis1, const b3Scalar& _angle)
 	{
+		b3Vector3 axis = axis1;
+		axis.safeNormalize();
+		
 		b3Scalar d = axis.length();
 		b3Assert(d != b3Scalar(0.0));
 		if (d < B3_EPSILON)