Allow max_relative_target to be a float

[jstnhuang]

What this does

This PR updates the robot configurations so that max_relative_target can be a float instead of just an integer.

This PR also makes a couple of small related changes:

• Updates the documentation for max_relative_target. Currently, it states that it can be a list of numbers, but the implementation of ensure_safe_goal_position only allows max_relative_target to be a scalar or a dictionary.
• Removes max_relative_target from Stretch3RobotConfig. Stretch3Robot does not use max_relative_target.
• Fixed an incorrect type annotation in ensure_safe_goal_position: dict[float] -> dict[str, float].

Fixes #1716 and fixes #1527.

How it was tested

I have an SO-101 and tested this change by running lerobot-teleoperate under 4 conditions:

• --robot.max_relative_target 10
  • Manually verified that teleop worked with goal clamping applied to all joints when moved quickly
• --robot.max_relative_target 10.5
  • Manually verified that teleop worked with goal clamping applied to all joints when moved quickly
• --robot.max_relative_target '{"shoulder_pan":20.0,"shoulder_lift":20.0,"elbow_flex":20.0,"wrist_flex":20.0,"wrist_roll":20.0,"gripper":5.0}'
  • Manually verified that teleop worked, and goal clamping only applied to the gripper
• And without --robot.max_relative_target
  • Manually verified that teleop worked with no goal clamping

I ran all the existing tests. All tests passed except for policy-related tests which I believe are unrelated environment setup issues:

FAILED tests/envs/test_envs.py::test_factory[aloha] - mujoco.FatalError: Default framebuffer is not complete, error 0x0 FAILED tests/policies/test_policies.py::test_policy[lerobot/aloha_sim_insertion_human-aloha-env_kwargs4-act-policy_kwargs4] - mujoco.FatalError: Default framebuffer is not complete, error 0x0 FAILED tests/policies/test_policies.py::test_policy[lerobot/aloha_sim_insertion_scripted-aloha-env_kwargs5-act-policy_kwargs5] - mujoco.FatalError: Default framebuffer is not complete, error 0x0 FAILED tests/policies/test_policies.py::test_policy[lerobot/aloha_sim_insertion_human-aloha-env_kwargs6-diffusion-policy_kwargs6] - mujoco.FatalError: Default framebuffer is not complete, error 0x0 FAILED tests/policies/test_policies.py::test_policy[lerobot/aloha_sim_transfer_cube_human-aloha-env_kwargs7-act-policy_kwargs7] - mujoco.FatalError: Default framebuffer is not complete, error 0x0 FAILED tests/policies/test_policies.py::test_policy[lerobot/aloha_sim_transfer_cube_scripted-aloha-env_kwargs8-act-policy_kwargs8] - mujoco.FatalError: Default framebuffer is not complete, error 0x0 6 failed, 816 passed, 12 skipped, 66 warnings in 193.68s (0:03:13)

How to checkout & try? (for the reviewer)

I tested this using a basic teleop command:

lerobot-teleoperate \
    --robot.type=so101_follower \
    --robot.port=$FOLLOWER \
    --robot.id=follower \
    --robot.cameras="$CAMERA_CONFIG" \
    --robot.max_relative_target 10 \
    --teleop.type=so101_leader \
    --teleop.port=$LEADER \
    --teleop.id=leader \
    --display_data=true

Where --robot.max_relative_target was set to one of the above.

[pkooij]

LGTM, left only one small comment

[pkooij]

Thanks!