Working CarRacing environment wrapper

[zlw21gxy]

I try to run the default car racing, but the result is weird,I follow the process in readme totally.May I ask why? but if I read my own wrapper for car racing, everything seems fine Maybe I misunderstand some details of the latest code... (ubuntu16.04 python3.5 tensorflow1.12.0)

[zlw21gxy]

def breakout(config, params):
  action_repeat = params.get('action_repeat', REPEATE)
  max_length = EPISODE_LEN // action_repeat
  state_components = ['reward']
  env_ctor = functools.partial(
      _dm_control_env_gym_atari, action_repeat, max_length, 'Breakout-v0')
  return Task('breakout', env_ctor, max_length, state_components)

def car_race(config, params):
  action_repeat = params.get('action_repeat', REPEATE)
  max_length = EPISODE_LEN // action_repeat
  state_components = ['reward']
  env_ctor = functools.partial(
      _dm_control_env_gym_atari, action_repeat, max_length, 'CarRacing-v0')
  return Task('car_race', env_ctor, max_length, state_components)

class DeepMindWrapper_gym_atari(object):
  """Wraps a Gym environment into an interface for downstream process"""

  metadata = {'render.modes': ['rgb_array']}
  reward_range = (-np.inf, np.inf)

  def __init__(self, env, render_size, camera_id=0):
    self._env = env
    self._render_size = render_size
    self._camera_id = camera_id

    self.observation_space = gym.spaces.Dict({'state':gym.spaces.Box(low=-1,high=1,shape=(1,))})

    self.action_space = gym.spaces.Box(low=-1,high=1,shape=(2,))

  def __getattr__(self, name):
    return getattr(self._env, name)

  def discrete_action(self, a):
    condition_list = [a>0.5, a>0.0, a>-0.5, True]
    choice_list = [0 ,1, 2 ,3]
    return np.select(condition_list, choice_list)

  def step(self, action):
    self._env.render()
    throttle = float(np.clip(action[1], 0, 1))
    brake = float(np.abs(np.clip(action[1], -1, 0)))
    steer = float(np.clip(action[0], -1, 1))
    action = (steer, throttle, brake)
    s_img, reward, done, info = self._env.step(action)
    self.img = cv2.resize(s_img, IMG_SIZE,interpolation=cv2.INTER_AREA)

    obs = {'state':np.array([0.0])}

    return obs, reward, False, {}     # done can be set to always False.

  def reset(self):
    s_img, info = self._env.reset()
    self.img = cv2.resize(s_img, IMG_SIZE, interpolation=cv2.INTER_AREA)

    return {'state': np.array([0.0])}

  def render(self, *args, **kwargs):
    if kwargs.get('mode', 'rgb_array') != 'rgb_array':
      raise ValueError("Only render mode 'rgb_array' is supported.")
    del args  # Unused
    del kwargs  # Unused
    return self.img

def _dm_control_env_gym_atari(action_repeat, max_length, env_name):
  import gym
  def env_ctor():
    env = gym.make(env_name)     # 'Breakout-v0'
    env = env.env                # 'remove the TimeLimit wrapper
    env = DeepMindWrapper_gym_atari(env, IMG_SIZE)
    env = control.wrappers.ActionRepeat(env, action_repeat)
    env = control.wrappers.LimitDuration(env, max_length)
    env = control.wrappers.PixelObservations(env, IMG_SIZE, np.uint8, 'image')
    env = control.wrappers.ConvertTo32Bit(env)
    return env
  env = control.wrappers.ExternalProcess(env_ctor)
  return env

This is my code of task, everything is fine, but when I use the latest version of planet which have car racing default, the agent behave weird

[danijar]

Thanks for your message! It's great to see that you got CarRacing to work. The existing environment wasn't tested well, since I just started to play around with it. The tested environments are the dm_control tasks from our paper.

If you like, it would be great if you would clean up your code a bit and then create a pull request to add it to the repository. This way, other people could easily train PlaNet on CarRacing and Breakout. Let me know if I can help with it.