A bug when using randomize_hand_positions

[zhaoyi11]

When using the augmentation randomize_hand_positions, there is a bug for the state of the physics.

robopianist/robopianist/suite/tasks/piano_with_shadow_hands.py

Lines 447 to 455 in d9cde23

	def _randomize_initial_hand_positions(
	self, physics: mjcf.Physics, random_state: np.random.RandomState
	) -> None:
	"""Randomize the initial position of the hands."""
	if not self._randomize_hand_positions:
	return
	offset = random_state.uniform(low=-_POSITION_OFFSET, high=_POSITION_OFFSET)
	for hand in [self.right_hand, self.left_hand]:
	hand.shift_pose(physics, (0, offset, 0))

Specifically, the shift applied to the initial hand position will cumulate. To reproduce this bug, one can run by setting the task_arguments with randomize_hand_positions=True.

I fix this issue in a hacky way by recording the previously applied offset and adding the offset back before applying a new offset. So the _randomize_initial_hand_positions changes to:

    def _randomize_initial_hand_positions(
        self, physics: mjcf.Physics, random_state: np.random.RandomState
    ) -> None:
        """Randomize the initial position of the hands."""
        if not self._randomize_hand_positions:
            return

        offset = random_state.uniform(low=-_POSITION_OFFSET, high=_POSITION_OFFSET)
        for hand in [self.right_hand, self.left_hand]:
            hand.shift_pose(physics, (0, -self._last_offset, 0)) # to correct the offset.
            hand.shift_pose(physics, (0, offset, 0))
        self._last_offset = offset

Hope you have a better solution. Or I can open a PR.

Best,
Yi