panda_robotiq controller not working as expected

[akuramshin]

Problem: sending only a change in gripper position results in the entire robot changing its pose.

Steps to reproduce:

1. Run python mani_skill/examples/demo_manual_control.py -e StackCube-v1 -c pd_ee_delta_pose --enable-sapien-viewer
   This should open up the StackCube environment with the panda_robotiq agent. The demo_manual_control script allows the user to manually control the robot with every action command being a single environment step.

2. Send a command to only open/close the gripper by pressing the f/g keys on your keyboard (you might have to first press on the matplotlib window for it to register keyboard commands).

Since we are using the pd_ee_delta_pose controller and we are only sending a gripper action, the expected outcome is for all the joints to remain static except for the gripper. However, the entire robot moves along with the gripper.

In contrast, changing [this line (https://github.com/montrealrobotics/ManiSkill/blob/main/mani_skill/envs/tasks/tabletop/stack_cube.py#L24) to use the "panda" embodiment instead will show you the expected outcome.

[akuramshin]

Currently following up in the ManiSkill repo in a thread here: haosulab#636 (comment)

[akuramshin]

Fixed! The link offset for the robotiq base from the last panda link was not far enough. I originally took the offsets from the urdf found here: https://github.com/facebookresearch/fairo/blob/main/polymetis/polymetis/data/franka_panda_robotiq_85/panda_robotiq_85.urdf

Looks like my URDF was wrong. The robotiq gripper base was not offset by the right amount from the last panda link.

[akuramshin]

Its important to note that the URDF is not complete. It is worth investigating how to match our URDF closer to the DROID setup. Additionally, we need to investigate how the gripper pose in the DROID dataset is calculated (relative to which link?)