deepmimic fix: reference initialization wasn't spread over all frames

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

@@ -68,7 +68,7 @@ class HumanoidStablePD(object):
     self.resetPose()
     
   def resetPose(self):
-    print("resetPose with self._frameFraction=",self._frameFraction)
+    print("resetPose with self._frame=", self._frame, " and self._frameFraction=",self._frameFraction)
     pose = self.computePose(self._frameFraction)
     self.initializePose(self._poseInterpolator, self._sim_model, initBase=True)
     self.initializePose(self._poseInterpolator, self._kin_model, initBase=False)
@@ -126,7 +126,6 @@ class HumanoidStablePD(object):
     keyFrameDuration = self._mocap_data.KeyFrameDuraction()
     cycleTime = self.getCycleTime()
     #print("self._motion_data.NumFrames()=",self._mocap_data.NumFrames())
-    #print("cycleTime=",cycleTime)
     cycles = self.calcCycleCount(t, cycleTime)
     #print("cycles=",cycles)
     frameTime = t - cycles*cycleTime

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

@@ -64,8 +64,12 @@ class PyBulletDeepMimicEnv(Env):
               #self._humanoid.applyPDForces(taus)
               #self._pybullet_client.stepSimulation()
             time.sleep(timeStep)
-
+      #print("numframes = ", self._humanoid._mocap_data.NumFrames())
       startTime = random.randint(0,self._humanoid._mocap_data.NumFrames()-2)
+      rnrange = 1000
+      rn = random.randint(0,rnrange)
+      startTime = float(rn)/rnrange * self._humanoid.getCycleTime()
+
       self._humanoid.setSimTime(startTime)
       self._humanoid.resetPose()
       #this clears the contact points. Todo: add API to explicitly clear all contact points?

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

@@ -3,6 +3,7 @@ import learning.agent_builder as AgentBuilder
 import learning.tf_util as TFUtil
 from learning.rl_agent import RLAgent
 from pybullet_utils.logger import Logger
+import pybullet_data
 
 class RLWorld(object):
     def __init__(self, env, arg_parser):
@@ -79,7 +80,7 @@ class RLWorld(object):
                 if (len(model_files) > 0):
                     curr_model_file = model_files[i]
                     if curr_model_file != 'none':
-                        curr_agent.load_model(curr_model_file)
+                        curr_agent.load_model(pybullet_data.getDataPath()+"/"+curr_model_file)
 
             self.agents.append(curr_agent)
             Logger.print2('')
@@ -139,4 +140,4 @@ class RLWorld(object):
             assert (agent != None), 'Failed to build agent {:d} from: {}'.format(id, agent_file)
         
         return agent
-        
+        

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

@@ -30,7 +30,7 @@ def build_arg_parser(args):
 args = sys.argv[1:]
 arg_parser = build_arg_parser(args)
 
-render=True
+render=False#True
 env = PyBulletDeepMimicEnv (args,render)
  
 world = RLWorld(env, arg_parser)
@@ -57,7 +57,6 @@ with open(agent_files) as data_file:
     
     agent.set_enable_training(True)
     world.reset()
-
     while (world.env._pybullet_client.isConnected()):
 
       timeStep = 1./600.