Fix Agibot left-arm teleop defects and env_cfg_callback contract

[qianl-nv]

Summary

Fix three defects that make Agibot left-arm teleoperation unusable and crash documented env_cfg_callback usage.

Detailed description

• Add a robot-only reset event to AgibotEmbodiment: event_config was never set, so every joint reset to 0 and RMPFlow's lula model no longer described the actual robot. Robot-only rather than mdp.reset_scene_to_default, which would also reset every scene asset and trample task-level contracts such as disable_reset_pose.
• Fix the left arm's frame/body offset quaternions, written as (w,x,y,z) in (x,y,z,w) fields: read as a 180° flip, the arm ran away ~270 mm under a zero command after every reset (corrected: 9.6 mm) and could not be teleoperated.
• ArenaEnvBuilder accept an env_cfg_callback that must return a cfg. Update the doc and comments to state so.
• Harden test_place_upright_mug_multi to assert the env_ids/percentage plumbing at the moment of the write: it previously depended on marginal rotate-in-place physics that any change to the robot's reset behaviour flips.

Resubmitted for #1097 since the CI doesn't work from fork.

[greptile-apps]

Greptile Summary

The PR repairs Agibot left-arm reset and frame-offset behavior while clarifying the environment-config callback contract.

• Adds an Agibot-specific reset event that restores default robot joint state and targets without resetting unrelated scene assets.
• Corrects the left-arm frame-transformer and RMPFlow body-offset quaternions.
• Documents and enforces that env_cfg_callback returns an environment configuration.
• Revises the multi-environment upright-placement test to validate selection and percentage plumbing immediately after writes.

Confidence Score: 5/5

The PR appears safe to merge with no concrete blocking or non-blocking defects identified.

The reset configuration, paired left-arm offsets, callback contract updates, and targeted test changes are internally consistent, and no reachable changed-code failure was established.

Important Files Changed

Filename	Overview
isaaclab_arena/embodiments/agibot/agibot.py	Adds robot-only reset behavior and consistently corrects the left-arm controller and observation-frame offsets; no actionable defect was established.
isaaclab_arena/environments/arena_env_builder.py	Adds a fail-fast assertion for callbacks that violate the existing return-value contract.
isaaclab_arena/environments/isaaclab_arena_environment.py	Updates callback typing and documentation to accurately express the builder's return-value contract.
isaaclab_arena/tests/test_place_upright_task.py	Refocuses the multi-environment test on immediate write plumbing rather than subsequent physics evolution.
docs/pages/concepts/environment/environment_definition.rst	Corrects the callback example by explicitly returning the mutated configuration.

_{Reviews (1): Last reviewed commit: "Simplify agibot reset comments" | Re-trigger Greptile}

[arena-review-bot]

🟡 A third, divergent "reset robot to default" implementation

We now have three approaches to the same RMPFlow "reset joints to their default" need: Franka and Galbot reuse franka_stack_events.randomize_joint_by_gaussian_offset (std=0), terms/events.py has reset_all_articulation_joints, and this adds a fourth bespoke one. The two things that make this one different are worth pulling out: it's robot-only (so it doesn't trample task assets) and it also restores the joint position/velocity targets. That last part is the interesting bit — if RMPFlow needs the targets reset too, then Galbot's gaussian-offset reset (which doesn't touch targets) likely has the same latent runaway you're fixing here.

Could this live in the shared terms/events.py as a robot-only reset_joints_to_default (with the target restore) so Galbot/Franka and any future RMPFlow embodiment converge on one implementation instead of drifting? Happy to be wrong if the target reset is genuinely Agibot-specific.

[arena-review-bot]

🔵 Mismatched backticks

`joint_lift_body `` opens with one backtick and closes with two, so it renders wrong. Also worth escaping the leading underscore names if this is meant as RST.

Suggested change

	RMPFlow expects `joint_lift_body`` and ``joint_body_pitch`` at those default values
	RMPFlow expects ``joint_lift_body`` and ``joint_body_pitch`` at those default values

[arena-review-bot]

🤖 Isaac Lab-Arena Review Bot

Summary

This PR fixes three real Agibot left-arm teleop defects (a never-set event_config so RMPFlow no longer saw the true joint state, and two left-arm offset quaternions written in the wrong component order) and tightens the env_cfg_callback contract to require returning the config. The quaternion fix is backed by the author's own runaway-distance measurements (270 mm → 9.6 mm) and the callback change is documented and asserted — both look right. Overall a clean, well-scoped fix; the one thing worth a look is that the new reset helper is a third divergent take on a problem two other embodiments already solve differently.

Findings

🟡 Warning: embodiments/agibot/agibot.py:151 — reset_robot_to_default is a fourth bespoke "reset joints to default" implementation alongside Franka/Galbot's randomize_joint_by_gaussian_offset(std=0) and terms/events.py:reset_all_articulation_joints. The distinguishing behavior — robot-only + restoring joint targets — is exactly what the others lack, so Galbot's target-less reset may carry the same latent runaway. Consider hoisting a robot-only reset_joints_to_default into the shared terms/events.py so RMPFlow embodiments converge on one path.

🔵 Improvement: embodiments/agibot/agibot.py:172 — mismatched backticks in the attribute docstring ( `joint_lift_body `` opens with one, closes with two).

Test Coverage

test_place_upright_mug_multi is hardened to assert the env_ids/upright_percentage plumbing at write time instead of after marginal rotate-in-place physics — a good call. NUM_STEPS and step_zeros_and_call are still used by the other two tests, so no dead code is left behind. No test exercises the new Agibot reset path directly, but that path needs a live sim + RMPFlow and is hard to unit-test; the empirical teleop verification is a reasonable substitute for an alpha fix.

Verdict

Ship it (consider the shared-reset consolidation as a follow-up).