Update xarm6_robotiq to have a more natural rest position #929
Conversation
dwaitbhatt
force-pushed
the
xarm_updates
branch
from
0bd2f02 to
d4bd1eb
Compare
| @@ -141,13 +141,9 @@ def compute_dense_reward(self, obs: Any, action: torch.Tensor, info: Dict): | |||
| place_reward = 1 - torch.tanh(5 * obj_to_goal_dist) | |||
| reward += place_reward * is_grasped | |||
|
|
|||
| qvel_without_gripper = self.agent.robot.get_qvel() | |||
| if self.robot_uids == "xarm6_robotiq": | |||
| qvel_without_gripper = qvel_without_gripper[..., :-6] | |||
There was a problem hiding this comment.
actually how come we don't keep this part? We need the joint velocity without gripper since the task objective is to ensure the robot stays still while grasping the cube.
Or at minimum keep the panda one (for compatability, all our previous experiments/numbers are based on qvel without gripper). If using the entire qvel works fine for other robots then we can just keep the exception for panda and use entire qvel for reward for the other robots.
There was a problem hiding this comment.
My rationale was that at the final stage gripper qpos can't change anyway for grasping tasks like PickCube, since if the gripper opens a bit then you'll drop the object, and you can't close more either because the objects are rigid. Hence encouraging 0 gripper qvel also makes sense.
For now I'll keep it in for panda for compatibility. Updating the PR accordingly.
| 1.52969832e-04, | ||
| 1.20606723e00, | ||
| 1.66234924e-03, | ||
| 6.672368e-4, |
There was a problem hiding this comment.
the e-4 values are probably supposed to be 0 right?
There was a problem hiding this comment.
Yeah, I had just copied joint positions from the GUI once I got the gripper at the right location. Quickly tested with qpos precision of 2 decimal places, that looks okay too. Updating the PR.
Update xarm6_robotiq rest qpos for completing tasks in fewer steps
Clean up pick cube reward to remove robot specific checks
Following discussion in #823