Merge pull request #1085 from jietan/pullRequest

A simple trained DDPG agent for minitaur walking

examples/pybullet/gym/agents/actor_net.py

@@ -0,0 +1,21 @@
+"""An actor network."""
+import tensorflow as tf
+import sonnet as snt
+
+class ActorNetwork(snt.AbstractModule):
+  """An actor network as a sonnet Module."""
+
+  def __init__(self, layer_sizes, action_size, name='target_actor'):
+    super(ActorNetwork, self).__init__(name=name)
+    self._layer_sizes = layer_sizes
+    self._action_size = action_size
+
+  def _build(self, inputs):
+    state = inputs
+    for output_size in self._layer_sizes:
+      state = snt.Linear(output_size)(state)
+      state = tf.nn.relu(state)
+
+    action = tf.tanh(
+        snt.Linear(self._action_size, name='action')(state))
+    return action

examples/pybullet/gym/agents/simpleAgent.py

@@ -0,0 +1,46 @@
+"""Loads a DDPG agent without too much external dependencies
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import collections
+import numpy as np
+import tensorflow as tf
+
+import sonnet as snt
+from agents import actor_net
+
+class SimpleAgent():
+  def __init__(
+      self,
+      session,
+      ckpt_path,
+      actor_layer_size,
+      observation_size=(31,),
+      action_size=8,
+  ):
+    self._ckpt_path = ckpt_path
+    self._actor_layer_size = actor_layer_size
+    self._observation_size = observation_size
+    self._action_size = action_size
+    self._session = session
+    self._build()
+
+  def _build(self):
+    self._agent_net = actor_net.ActorNetwork(self._actor_layer_size, self._action_size)
+    self._obs = tf.placeholder(tf.float32, (31,))
+    with tf.name_scope('Act'):
+      batch_obs = snt.nest.pack_iterable_as(self._obs,
+                                            snt.nest.map(lambda x: tf.expand_dims(x, 0),
+                                                         snt.nest.flatten_iterable(self._obs)))
+      self._action = self._agent_net(batch_obs)
+      saver = tf.train.Saver()
+      saver.restore(
+          sess=self._session,
+          save_path=self._ckpt_path)
+
+  def __call__(self, observation):
+    out_action = self._session.run(self._action, feed_dict={self._obs: observation})
+    return out_action[0]

examples/pybullet/gym/data/agent/tf_graph_data/checkpoint

@@ -0,0 +1,2 @@
+model_checkpoint_path: "/cns/ij-d/home/jietan/persistent/minitaur/minitaur_vizier_3_153645653/Bullet/MinitaurSimEnv/28158/0003600000/agent/tf_graph_data/tf_graph_data.ckpt"
+all_model_checkpoint_paths: "/cns/ij-d/home/jietan/persistent/minitaur/minitaur_vizier_3_153645653/Bullet/MinitaurSimEnv/28158/0003600000/agent/tf_graph_data/tf_graph_data.ckpt"

examples/pybullet/gym/envs/bullet/minitaurGymEnv.py

@@ -0,0 +1,109 @@
+import math
+import gym
+from gym import spaces
+from gym.utils import seeding
+import numpy as np
+import time
+import pybullet as p
+import minitaur_new
+
+class MinitaurGymEnv(gym.Env):
+  metadata = {
+      'render.modes': ['human', 'rgb_array'],
+      'video.frames_per_second' : 50
+  }
+
+  def __init__(self,
+               urdfRoot="",
+               actionRepeat=1,
+               render=False):
+    self._timeStep = 0.01
+    self._urdfRoot = urdfRoot
+    self._actionRepeat = actionRepeat
+    self._observation = []
+    self._envStepCounter = 0
+    self._render = render
+    self._lastBasePosition = [0, 0, 0]
+    if self._render:
+      p.connect(p.GUI)
+    else:
+      p.connect(p.DIRECT)
+    self._seed()
+    self.reset()
+    observationDim = self._minitaur.getObservationDimension()
+    observation_high = np.array([np.finfo(np.float32).max] * observationDim)
+    actionDim = 8
+    action_high = np.array([1] * actionDim)
+    self.action_space = spaces.Box(-action_high, action_high)
+    self.observation_space = spaces.Box(-observation_high, observation_high)
+    self.viewer = None
+
+  def _reset(self):
+    p.resetSimulation()
+    p.setPhysicsEngineParameter(numSolverIterations=300)
+    p.setTimeStep(self._timeStep)
+    p.loadURDF("%splane.urdf" % self._urdfRoot)
+    p.setGravity(0,0,-10)
+    self._minitaur = minitaur_new.Minitaur(self._urdfRoot)
+    self._envStepCounter = 0
+    self._lastBasePosition = [0, 0, 0]
+    for i in range(100):
+      p.stepSimulation()
+    self._observation = self._minitaur.getObservation()
+    return self._observation
+
+  def __del__(self):
+    p.disconnect()
+
+  def _seed(self, seed=None):
+    self.np_random, seed = seeding.np_random(seed)
+    return [seed]
+
+  def _step(self, action):
+    if (self._render):
+      basePos = self._minitaur.getBasePosition()
+      p.resetDebugVisualizerCamera(1, 30, 40, basePos)
+
+    if len(action) != self._minitaur.getActionDimension():
+      raise ValueError("We expect {} continuous action not {}.".format(self._minitaur.getActionDimension(), len(action)))
+
+    for i in range(len(action)):
+      if not -1.01 <= action[i] <= 1.01:
+        raise ValueError("{}th action should be between -1 and 1 not {}.".format(i, action[i]))
+
+    realAction = self._minitaur.convertFromLegModel(action)
+    self._minitaur.applyAction(realAction)
+    for i in range(self._actionRepeat):
+      p.stepSimulation()
+      if self._render:
+        time.sleep(self._timeStep)
+      self._observation = self._minitaur.getObservation()
+      if self._termination():
+        break
+      self._envStepCounter += 1
+    reward = self._reward()
+    done = self._termination()
+    return np.array(self._observation), reward, done, {}
+
+  def _render(self, mode='human', close=False):
+      return
+
+  def is_fallen(self):
+    orientation = self._minitaur.getBaseOrientation()
+    rotMat = p.getMatrixFromQuaternion(orientation)
+    localUp = rotMat[6:]
+    return np.dot(np.asarray([0, 0, 1]), np.asarray(localUp)) < 0 or self._observation[-1] < 0.1
+
+  def _termination(self):
+    return self.is_fallen()
+
+  def _reward(self):
+    currentBasePosition = self._minitaur.getBasePosition()
+    forward_reward = currentBasePosition[0] - self._lastBasePosition[0]
+    self._lastBasePosition = currentBasePosition
+
+    energyWeight = 0.001
+    energy = np.abs(np.dot(self._minitaur.getMotorTorques(), self._minitaur.getMotorVelocities())) * self._timeStep
+    energy_reward = energyWeight * energy
+    reward = forward_reward - energy_reward
+    return reward

examples/pybullet/gym/envs/bullet/minitaur_new.py

@@ -0,0 +1,166 @@
+import pybullet as p
+import numpy as np
+import copy
+import math
+
+class Minitaur:
+  def __init__(self, urdfRootPath=''):
+    self.urdfRootPath = urdfRootPath
+    self.reset()
+
+  def buildJointNameToIdDict(self):
+    nJoints = p.getNumJoints(self.quadruped)
+    self.jointNameToId = {}
+    for i in range(nJoints):
+      jointInfo = p.getJointInfo(self.quadruped, i)
+      self.jointNameToId[jointInfo[1].decode('UTF-8')] = jointInfo[0]
+    self.resetPose()
+
+  def buildMotorIdList(self):
+    self.motorIdList.append(self.jointNameToId['motor_front_leftL_joint'])
+    self.motorIdList.append(self.jointNameToId['motor_front_leftR_joint'])
+    self.motorIdList.append(self.jointNameToId['motor_back_leftL_joint'])
+    self.motorIdList.append(self.jointNameToId['motor_back_leftR_joint'])
+    self.motorIdList.append(self.jointNameToId['motor_front_rightL_joint'])
+    self.motorIdList.append(self.jointNameToId['motor_front_rightR_joint'])
+    self.motorIdList.append(self.jointNameToId['motor_back_rightL_joint'])
+    self.motorIdList.append(self.jointNameToId['motor_back_rightR_joint'])
+
+  def reset(self):
+#    self.quadruped = p.loadURDF("%squadruped/minitaur.urdf" % self.urdfRootPath, [0,0,.2], flags=p.URDF_USE_SELF_COLLISION)
+    self.quadruped = p.loadURDF("%squadruped/minitaur.urdf" % self.urdfRootPath, [0,0,.2])
+    self.kp = 1
+    self.kd = 1
+    self.maxForce = 3.5
+    self.nMotors = 8
+    self.motorIdList = []
+    self.motorDir = [-1, -1, -1, -1, 1, 1, 1, 1]
+    self.buildJointNameToIdDict()
+    self.buildMotorIdList()
+
+
+  def setMotorAngleById(self, motorId, desiredAngle):
+    p.setJointMotorControl2(bodyIndex=self.quadruped, jointIndex=motorId, controlMode=p.POSITION_CONTROL, targetPosition=desiredAngle, positionGain=self.kp, velocityGain=self.kd, force=self.maxForce)
+
+  def setMotorAngleByName(self, motorName, desiredAngle):
+    self.setMotorAngleById(self.jointNameToId[motorName], desiredAngle)
+
+  def resetPose(self):
+    kneeFrictionForce = 0
+    halfpi = 1.57079632679
+    kneeangle = -2.1834 #halfpi - acos(upper_leg_length / lower_leg_length)
+
+    #left front leg
+    p.resetJointState(self.quadruped,self.jointNameToId['motor_front_leftL_joint'],self.motorDir[0]*halfpi)
+    p.resetJointState(self.quadruped,self.jointNameToId['knee_front_leftL_link'],self.motorDir[0]*kneeangle)
+    p.resetJointState(self.quadruped,self.jointNameToId['motor_front_leftR_joint'],self.motorDir[1]*halfpi)
+    p.resetJointState(self.quadruped,self.jointNameToId['knee_front_leftR_link'],self.motorDir[1]*kneeangle)
+    p.createConstraint(self.quadruped,self.jointNameToId['knee_front_leftR_link'],self.quadruped,self.jointNameToId['knee_front_leftL_link'],p.JOINT_POINT2POINT,[0,0,0],[0,0.005,0.2],[0,0.01,0.2])
+    self.setMotorAngleByName('motor_front_leftL_joint', self.motorDir[0]*halfpi)
+    self.setMotorAngleByName('motor_front_leftR_joint', self.motorDir[1]*halfpi)
+    p.setJointMotorControl2(bodyIndex=self.quadruped,jointIndex=self.jointNameToId['knee_front_leftL_link'],controlMode=p.VELOCITY_CONTROL,targetVelocity=0,force=kneeFrictionForce)
+    p.setJointMotorControl2(bodyIndex=self.quadruped,jointIndex=self.jointNameToId['knee_front_leftR_link'],controlMode=p.VELOCITY_CONTROL,targetVelocity=0,force=kneeFrictionForce)
+
+    #left back leg
+    p.resetJointState(self.quadruped,self.jointNameToId['motor_back_leftL_joint'],self.motorDir[2]*halfpi)
+    p.resetJointState(self.quadruped,self.jointNameToId['knee_back_leftL_link'],self.motorDir[2]*kneeangle)
+    p.resetJointState(self.quadruped,self.jointNameToId['motor_back_leftR_joint'],self.motorDir[3]*halfpi)
+    p.resetJointState(self.quadruped,self.jointNameToId['knee_back_leftR_link'],self.motorDir[3]*kneeangle)
+    p.createConstraint(self.quadruped,self.jointNameToId['knee_back_leftR_link'],self.quadruped,self.jointNameToId['knee_back_leftL_link'],p.JOINT_POINT2POINT,[0,0,0],[0,0.005,0.2],[0,0.01,0.2])
+    self.setMotorAngleByName('motor_back_leftL_joint',self.motorDir[2]*halfpi)
+    self.setMotorAngleByName('motor_back_leftR_joint',self.motorDir[3]*halfpi)
+    p.setJointMotorControl2(bodyIndex=self.quadruped,jointIndex=self.jointNameToId['knee_back_leftL_link'],controlMode=p.VELOCITY_CONTROL,targetVelocity=0,force=kneeFrictionForce)
+    p.setJointMotorControl2(bodyIndex=self.quadruped,jointIndex=self.jointNameToId['knee_back_leftR_link'],controlMode=p.VELOCITY_CONTROL,targetVelocity=0,force=kneeFrictionForce)
+ 
+    #right front leg
+    p.resetJointState(self.quadruped,self.jointNameToId['motor_front_rightL_joint'],self.motorDir[4]*halfpi)
+    p.resetJointState(self.quadruped,self.jointNameToId['knee_front_rightL_link'],self.motorDir[4]*kneeangle)
+    p.resetJointState(self.quadruped,self.jointNameToId['motor_front_rightR_joint'],self.motorDir[5]*halfpi)
+    p.resetJointState(self.quadruped,self.jointNameToId['knee_front_rightR_link'],self.motorDir[5]*kneeangle)
+    p.createConstraint(self.quadruped,self.jointNameToId['knee_front_rightR_link'],self.quadruped,self.jointNameToId['knee_front_rightL_link'],p.JOINT_POINT2POINT,[0,0,0],[0,0.005,0.2],[0,0.01,0.2])
+    self.setMotorAngleByName('motor_front_rightL_joint',self.motorDir[4]*halfpi)
+    self.setMotorAngleByName('motor_front_rightR_joint',self.motorDir[5]*halfpi)
+    p.setJointMotorControl2(bodyIndex=self.quadruped,jointIndex=self.jointNameToId['knee_front_rightL_link'],controlMode=p.VELOCITY_CONTROL,targetVelocity=0,force=kneeFrictionForce)
+    p.setJointMotorControl2(bodyIndex=self.quadruped,jointIndex=self.jointNameToId['knee_front_rightR_link'],controlMode=p.VELOCITY_CONTROL,targetVelocity=0,force=kneeFrictionForce)
+ 
+
+    #right back leg
+    p.resetJointState(self.quadruped,self.jointNameToId['motor_back_rightL_joint'],self.motorDir[6]*halfpi)
+    p.resetJointState(self.quadruped,self.jointNameToId['knee_back_rightL_link'],self.motorDir[6]*kneeangle)
+    p.resetJointState(self.quadruped,self.jointNameToId['motor_back_rightR_joint'],self.motorDir[7]*halfpi)
+    p.resetJointState(self.quadruped,self.jointNameToId['knee_back_rightR_link'],self.motorDir[7]*kneeangle)
+    p.createConstraint(self.quadruped,self.jointNameToId['knee_back_rightR_link'],self.quadruped,self.jointNameToId['knee_back_rightL_link'],p.JOINT_POINT2POINT,[0,0,0],[0,0.005,0.2],[0,0.01,0.2])
+    self.setMotorAngleByName('motor_back_rightL_joint',self.motorDir[6]*halfpi)
+    self.setMotorAngleByName('motor_back_rightR_joint',self.motorDir[7]*halfpi)
+    p.setJointMotorControl2(bodyIndex=self.quadruped,jointIndex=self.jointNameToId['knee_back_rightL_link'],controlMode=p.VELOCITY_CONTROL,targetVelocity=0,force=kneeFrictionForce)
+    p.setJointMotorControl2(bodyIndex=self.quadruped,jointIndex=self.jointNameToId['knee_back_rightR_link'],controlMode=p.VELOCITY_CONTROL,targetVelocity=0,force=kneeFrictionForce)
+
+
+  def getBasePosition(self):
+    position, orientation = p.getBasePositionAndOrientation(self.quadruped)
+    return position
+
+  def getBaseOrientation(self):
+    position, orientation = p.getBasePositionAndOrientation(self.quadruped)
+    return orientation
+
+  def getActionDimension(self):
+    return self.nMotors
+
+  def getObservationDimension(self):
+    return len(self.getObservation())
+
+  def getObservation(self):
+    observation = []
+    observation.extend(self.getMotorAngles().tolist())
+    observation.extend(self.getMotorVelocities().tolist())
+    observation.extend(self.getMotorTorques().tolist())
+    observation.extend(list(self.getBaseOrientation()))
+    observation.extend(list(self.getBasePosition()))
+    return observation
+
+
+  def applyAction(self, motorCommands):
+    motorCommandsWithDir = np.multiply(motorCommands, self.motorDir)
+#    print('action: {}'.format(motorCommands))
+#    print('motor: {}'.format(motorCommandsWithDir))
+    for i in range(self.nMotors):
+      self.setMotorAngleById(self.motorIdList[i], motorCommandsWithDir[i])
+
+  def getMotorAngles(self):
+    motorAngles = []
+    for i in range(self.nMotors):
+      jointState = p.getJointState(self.quadruped, self.motorIdList[i])
+      motorAngles.append(jointState[0])
+    motorAngles = np.multiply(motorAngles, self.motorDir)
+    return motorAngles
+
+  def getMotorVelocities(self):
+    motorVelocities = []
+    for i in range(self.nMotors):
+      jointState = p.getJointState(self.quadruped, self.motorIdList[i])
+      motorVelocities.append(jointState[1])
+    motorVelocities = np.multiply(motorVelocities, self.motorDir)
+    return motorVelocities
+
+  def getMotorTorques(self):
+    motorTorques = []
+    for i in range(self.nMotors):
+      jointState = p.getJointState(self.quadruped, self.motorIdList[i])
+      motorTorques.append(jointState[3])
+    motorTorques = np.multiply(motorTorques, self.motorDir)
+    return motorTorques
+
+  def convertFromLegModel(self, actions):
+    motorAngle = copy.deepcopy(actions)
+    scaleForSingularity = 1
+    offsetForSingularity = 0.5
+    motorAngle[0] = math.pi + math.pi / 4 * actions[0] - scaleForSingularity * math.pi / 4 * (actions[4] + 1 + offsetForSingularity)
+    motorAngle[1] = math.pi - math.pi / 4 * actions[0] - scaleForSingularity * math.pi / 4 * (actions[4] + 1 + offsetForSingularity)
+    motorAngle[2] = math.pi + math.pi / 4 * actions[1] - scaleForSingularity * math.pi / 4 * (actions[5] + 1 + offsetForSingularity)
+    motorAngle[3] = math.pi - math.pi / 4 * actions[1] - scaleForSingularity * math.pi / 4 * (actions[5] + 1 + offsetForSingularity)
+    motorAngle[4] = math.pi - math.pi / 4 * actions[2] - scaleForSingularity * math.pi / 4 * (actions[6] + 1 + offsetForSingularity)
+    motorAngle[5] = math.pi + math.pi / 4 * actions[2] - scaleForSingularity * math.pi / 4 * (actions[6] + 1 + offsetForSingularity)
+    motorAngle[6] = math.pi - math.pi / 4 * actions[3] - scaleForSingularity * math.pi / 4 * (actions[7] + 1 + offsetForSingularity)
+    motorAngle[7] = math.pi + math.pi / 4 * actions[3] - scaleForSingularity * math.pi / 4 * (actions[7] + 1 + offsetForSingularity)
+    return motorAngle

examples/pybullet/gym/minitaurGymEnvTest.py

@@ -0,0 +1,85 @@
+'''
+A test for minitaurGymEnv
+'''
+
+import gym
+import numpy as np
+import math
+
+import numpy as np
+import tensorflow as tf
+
+from envs.bullet.minitaurGymEnv import MinitaurGymEnv
+from agents import simpleAgent
+
+def testSinePolicy():
+  """Tests sine policy
+    """
+  np.random.seed(47)
+
+  environment = MinitaurGymEnv(render=True)
+  sum_reward = 0
+  steps = 1000
+  amplitude1Bound = 0.5
+  amplitude2Bound = 0.15
+  speed = 40
+
+  for stepCounter in range(steps):
+    t = float(stepCounter) * environment._timeStep
+
+    if (t < 1):
+      amplitude1 = 0
+      amplitude2 = 0
+    else:
+      amplitude1 = amplitude1Bound
+      amplitude2 = amplitude2Bound
+    a1 = math.sin(t*speed)*amplitude1
+    a2 = math.sin(t*speed+3.14)*amplitude1
+    a3 = math.sin(t*speed)*amplitude2
+    a4 = math.sin(t*speed+3.14)*amplitude2
+
+    action = [a1, a2, a2, a1, a3, a4, a4, a3]
+
+    state, reward, done, info = environment.step(action)
+    sum_reward += reward
+    if done:
+      environment.reset()
+      print("sum reward: ", sum_reward)
+
+
+def testDDPGPolicy():
+  """Tests sine policy
+    """
+  environment = MinitaurGymEnv(render=True)
+  sum_reward = 0
+  steps = 1000
+  ckpt_path = 'data/agent/tf_graph_data/tf_graph_data.ckpt'
+  observation_shape = (31,)
+  action_size = 8
+  actor_layer_sizes = (100, 181)
+  n_steps = 0
+  tf.reset_default_graph()
+  with tf.Session() as session:
+    agent = simpleAgent.SimpleAgent(session, ckpt_path,
+                                    actor_layer_sizes,
+                                    observation_size=observation_shape,
+                                    action_size=action_size)
+    state = environment.reset()
+    action = agent(state)
+    for _ in range(steps):
+      n_steps += 1
+      state, reward, done, info = environment.step(action)
+      action = agent(state)
+      sum_reward += reward
+      if done:
+        environment.reset()
+        n_steps += 1
+        print("total reward: ", sum_reward)
+        print("total steps:  ", n_steps)
+        sum_reward = 0
+        n_steps = 0
+        return
+
+
+testDDPGPolicy()
+#testSinePolicy()