Class Humanoid

dm_control/suite/humanoid.py:132–207

A humanoid task.

class Humanoid(base.Task):
  """A humanoid task."""

  def __init__(self, move_speed, pure_state, random=None):
    """Initializes an instance of `Humanoid`.

    Args:
      move_speed: A float. If this value is zero, reward is given simply for
        standing up. Otherwise this specifies a target horizontal velocity for
        the walking task.
      pure_state: A bool. Whether the observations consist of the pure MuJoCo
        state or includes some useful features thereof.
      random: Optional, either a `numpy.random.RandomState` instance, an
        integer seed for creating a new `RandomState`, or None to select a seed
        automatically (default).
    """
    self._move_speed = move_speed
    self._pure_state = pure_state
    super().__init__(random=random)

  def initialize_episode(self, physics):
    """Sets the state of the environment at the start of each episode.

    Args:
      physics: An instance of `Physics`.

    """
    # Find a collision-free random initial configuration.
    penetrating = True
    while penetrating:
      randomizers.randomize_limited_and_rotational_joints(physics, self.random)
      # Check for collisions.
      physics.after_reset()
      penetrating = physics.data.ncon > 0
    super().initialize_episode(physics)

  def get_observation(self, physics):
    """Returns either the pure state or a set of egocentric features."""
    obs = collections.OrderedDict()
    if self._pure_state:
      obs['position'] = physics.position()
      obs['velocity'] = physics.velocity()
    else:
      obs['joint_angles'] = physics.joint_angles()
      obs['head_height'] = physics.head_height()
      obs['extremities'] = physics.extremities()
      obs['torso_vertical'] = physics.torso_vertical_orientation()
      obs['com_velocity'] = physics.center_of_mass_velocity()
      obs['velocity'] = physics.velocity()
    return obs

  def get_reward(self, physics):
    """Returns a reward to the agent."""
    standing = rewards.tolerance(physics.head_height(),
                                 bounds=(_STAND_HEIGHT, float('inf')),
                                 margin=_STAND_HEIGHT/4)
    upright = rewards.tolerance(physics.torso_upright(),
                                bounds=(0.9, float('inf')), sigmoid='linear',