Added new teleoperation script + PushCubeLoop env + small refactoring of the environments.

[michel-aractingi]

What does this PR do:

• Added new teleoperation script with LeRobot code in examples/teleoperate_leader_lerobot.py
• Added new PushCubeLoop-v0 env in push_cube_loop_env.py to mimic the RL task currently used in LeRobot.
• Added visualizations to PickPlace-v0 and PushCube-v0 in order to visualize the target region.
• Added indexing to the cube and arm poses based on their body name in the sim in all envs rather than constant indices.
• Added control decimation, i.e. multiple simulation steps per control step.
• Improved the contact between gripper and cube by modifying the collision params in the description.

Bug fixes:

• Fixed bug in teleoperation by sending control through mujoco rather than modifying joint in data.
• Fixed index issue as after refactor: update MuJoCo XML and STL files, remove outdated ones #33 the order of the robot and cubes was inverted so the indices of the joint and cube were completely wrong.
• Fixed robot orientation. The new robot description added in refactor: update MuJoCo XML and STL files, remove outdated ones #33 was rotated by 90 degrees so the robot was facing the cube in all environments.

How to checkout and test:

In order to test the teleoperation you should:

1. Get the start_pos from the calibration done by LeRobot and change the values in examples/teleoperate_leader_lerobot.py. (can be found in .cache/calibration/koch/main_leader.json)
2. Run (use mjpython for mac):

python examples/teleoperate_leader_lerobot.py --device /dev/tty.usbmodem585A0078211 --env-name PushCubeLoop-v0

depending on your device and environment you'd like to test.

TODO

Learning:

• Validate action space: instead of absolute joint angle commands add a delta joint angle commands.
• Validate reward function: debug reward function by trying to learn the tasks with LeRobot's tdmpc
• Add sparse reward functions.