def _compute_reward(self, goal, obs):
goal_pos, = goal
qpos, = obs
if at_goal(self._gripper_pos(qpos), goal_pos, self._geofence):
return 1
elif self._neg_reward:
return -.0001
else:
return 0
def _compute_reward(self, goal, obs):
pos = obs[0]
if at_goal(pos, goal, self._geofence):
return 1
elif escaped(pos, self._world_upper_bound, self._world_lower_bound):
return -1
elif self._neg_reward:
return -0.01
else:
return 0
def _compute_terminal(self, goal, obs):
goal, = goal
qpos, = obs
return at_goal(self._gripper_pos(qpos), goal, self._geofence)
def _achieved_goal(self, goal, obs):
goal_pos, (should_lift, ) = goal
qpos, (fingers_touching, block_lifted) = obs
_at_goal = at_goal(self._gripper_pos(qpos), goal_pos, self._geofence)
return _at_goal and should_lift == (block_lifted and fingers_touching)
def at_goal(self, goal, new_obs):
without_goal = new_obs[:-2]
position_orientation = without_goal[-4:]
position = position_orientation[:2]
return at_goal(position, goal, self._geofence)
def _compute_terminal(self, goal, obs):
goal, = goal
pos = obs[0]
return at_goal(pos, goal, self._geofence)