Adds support for gear insertion with Flexiv Rizon 4s

[shauryadNv]

Description

Added support to train a gear assembly/insertion policy for the Flexiv Rizon 4s robot. Defined new envs/tasks for this.
Added required robot config for the Flexiv Rizon 4s with Grav gripper.
Added documentation for a tutorial on training a gear assembly policy with Flexiv Rizon 4s.

Type of change

• New feature (non-breaking change which adds functionality)
• Documentation update

Checklist

• I have read and understood the contribution guidelines
• I have run the pre-commit checks with ./isaaclab.sh --format
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• I have updated the changelog and the corresponding version in the extension's config/extension.toml file
• I have added my name to the CONTRIBUTORS.md or my name already exists there

[AntoineRichard]

Thanks for the work @shauryadNv ! If this PR is not ready for review would you mind putting in draft mode?

[AntoineRichard]

I'm switching this PR to draft, please make sure you address all of the bots comments before moving it to "ready" again. Most of the comments are relevant.

[isaaclab-review-bot]

🤖 Follow-up Review (4dfee2d)

26 new commits since initial review. Good progress — here is what changed and what still needs attention.

✅ Fixed

• Simulation frequency reverted — decimation=4 and dt=1/120 restored in base __post_init__. UR10e environments no longer get unexpected physics timestep changes.
• eval.py removed — replaced with play_hardcoded.py, which resolves the C901 complexity pre-commit failure. The new script is well-documented with clear observation override constants and old-checkpoint conversion support.

🔴 Still Critical

• New reward terms in base RewardsCfg still break UR10e. end_effector_base_keypoint_tracking and end_effector_base_keypoint_tracking_exp are defined in RewardsCfg (shared base) but require end_effector_body_name, grasp_rot_offset, and gear_offsets_grasp in their params. Only the Rizon4s __post_init__ (line 410-418 in joint_pos_env_cfg.py) populates these — the UR10e configs do not. The keypoint_ee_gear_error.__init__ does cfg.params["end_effector_body_name"] which will KeyError for all UR10e environments. Fix: Either move these reward terms to a Rizon4s-specific RewardsCfg, or have the UR10e configs also populate these params.

🟡 Still Open

• Observation noise doubled — gear_shaft_pos noise changed from ±5mm to ±10mm in the base class, affecting UR10e too.
• is_close variable semantically inverted — is_close = (mean_kp_error > threshold) is True when the gripper is FAR. With threshold=0.0 it is always True so no functional bug currently, but misleading.
• import carb inside __call__ — runs every step in the hot path. Move to module level.
• ~90% code duplication between keypoint_ee_gear_error and keypoint_ee_gear_error_exp. Extract shared logic (init, gear lookup, grasp transform, weight scaling) into a base class or helper.
• Commented-out code — # action = RewTerm(func=mdp.action_l2, weight=-5.0e-06) in RewardsCfg. Remove or uncomment.

🔵 Minor / Docs

• Doc mismatch — gear_assembly_policy.rst still shows IdealPDActuatorCfg; code uses ImplicitActuatorCfg.
• Missing CHANGELOG entries for new robot config, environments, reward functions, and quaternion bug fix.

❌ CI

• pre-commit still failing — ruff format needs 1 file reformatted. Run pre-commit run --all-files locally.

New Files (Quick Assessment)

• play_hardcoded.py — Clean script. The _convert_old_checkpoint function is a nice addition for backward compat. Observation slice constants (_SHAFT_POS_SLICE, _SHAFT_QUAT_SLICE) are hardcoded to indices 14-21 which couples tightly to the obs layout — a brief comment noting the expected format would help.
• hardcoded_inference_env_cfg.py — Clean config inheriting from ROS inference and zeroing out randomization. The GEAR_BASE_ROT quaternion (0.0, 0.0, 0.70711, -0.70711) uses (x,y,z,w) format per the comment — correct and consistent.
• ResetSampledQuaternionNoiseModel — Well-designed noise model using quaternion multiplication instead of additive noise. Currently configured with zero ranges so no effect, but good infrastructure for future use.