Integrate Droid lightwheel kitchen example with robot placement

[qianl-nv]

Summary

Add Droid kitchen env yaml with robot placement

Detailed description

• Add droid_pick_and_place_lightwheel_kitchen.yaml with Droid-on-floor/counter relations in the Lightwheel kitchen scene.
• Use stand footprint for Droid placement bounds and write robot + stand roots on reset via layout_pose_to_scene_writes, adapted to the unified PlacementAsset/EmbodimentBase pose-reset path and main #945 stand-height API.
• Extend placement-event and relation-solver tests for Droid stand reset, bbox sizing.

Command to run (start pi server first):
/isaac-sim/python.sh isaaclab_arena/evaluation/policy_runner.py --viz kit --policy_type isaaclab_arena_openpi.policy.pi0_remote_policy.Pi0RemotePolicy --num_episodes 2 --enable_cameras --record_viewport_video --output_base_dir output/kitchen_task --env_graph_spec_yaml isaaclab_arena_environments/droid_pick_and_place_lightwheel_kitchen.yaml

Intermediate results:

• On base:

• Switch to stand bbox without stand height fix:

• Stand bbox + stand height fix:

TODO: Enable mesh-based placement for background and embodiment

[qianl-nv]

self review 1

[greptile-apps]

Greptile Summary

This PR integrates the Droid robot into the Lightwheel kitchen environment by adding a new YAML spec and extending the placement/reset pipeline to write both the robot root and its separate stand prim on each reset. The key mechanical change is layout_pose_to_scene_writes on DroidEmbodimentBase, which lifts the solver layout pose by the stand-height offset and composes the stand's local root offset via compose_poses, plus a new _write_scene_root_pose_to_sim helper that dispatches between write_root_pose_to_sim (articulations/rigid bodies) and set_world_poses (static prims like the stand).

• droid.py: Adds get_bounding_box (stand footprint as placement proxy), _stand_pose_from_robot_pose, and layout_pose_to_scene_writes; refactors stand init_state to use the shared _STAND_ROOT_OFFSET_IN_ROBOT_FRAME constant instead of the robot pose directly.
• placement_events.py: Extracts _write_scene_root_pose_to_sim to handle both physics and non-physics scene assets polymorphically; the write_layout_to_sim loop now uses it.
• Test files: Expanded with stand reset alignment, bbox sizing, and relation-solver stand-write tests; the kitchen reset alignment check contains a broken API call (get_world_poses()[0].torch) that will raise AttributeError at runtime, so the alignment invariant is never actually verified.

Confidence Score: 3/5

The production placement code is logically correct, but the new kitchen reset alignment test will always crash before making any assertion, so the robot-z vs stand-z invariant it is meant to guard is never verified.

The core stand-height offset and dual-prim reset logic is well-reasoned and consistent with the existing codebase. The broken .torch access in _check_kitchen_reset_alignment means the test cannot pass on any run, leaving the most important integration-level check non-functional. The stand mock issue in test_placement_events.py further means the set_world_poses dispatch path for the stand is untested, so a regression there would be silent.

test_droid_stand_height.py (_check_kitchen_reset_alignment) and test_placement_events.py (test_solve_and_place_objects_writes_droid_robot_and_stand) need attention before the test suite can provide meaningful coverage of the new reset path.

Important Files Changed

Filename	Overview
isaaclab_arena/embodiments/droid/droid.py	Adds get_bounding_box (stand footprint proxy), _stand_pose_from_robot_pose, and layout_pose_to_scene_writes; refactors stand init_state to use the shared _STAND_ROOT_OFFSET_IN_ROBOT_FRAME constant. Logic is sound but the stand scene-entity name is hardcoded.
isaaclab_arena/relations/placement_events.py	Extracts _write_scene_root_pose_to_sim to handle both articulation (write_root_pose_to_sim) and static-mesh (set_world_poses) assets. Dispatch logic is correct; the unreachable sentinel uses assert False which is silently dropped under -O.
isaaclab_arena/tests/test_droid_stand_height.py	Expanded test suite for stand scaling, pose wiring, bbox, and kitchen reset alignment; _check_kitchen_reset_alignment accesses get_world_poses()[0].torch which is not a valid attribute on torch.Tensor and will raise AttributeError at runtime.
isaaclab_arena/tests/test_placement_events.py	Adds Droid robot+stand placement test and missing-asset assertion test; the Droid test uses MagicMock which auto-creates write_root_pose_to_sim, hiding that the stand should use set_world_poses in production.
isaaclab_arena_environments/droid_pick_and_place_lightwheel_kitchen.yaml	New environment spec for Droid in the Lightwheel kitchen scene; relations (on floor, next_to counter, rotate, object placement) are well-formed and consistent with the arena env graph format.

_{Reviews (1): Last reviewed commit: "Consolidate Droid stand and placement te..." | Re-trigger Greptile}

[arena-review-bot]

🤖 Isaac Lab-Arena Review Bot

Summary

This PR adds a Droid + Lightwheel-kitchen environment yaml and, to support it, makes the Droid embodiment place both the robot root and its separate stand on reset (via the layout_pose_to_scene_writes compound-asset hook) and use the stand footprint as its placement bbox. The scene-write plumbing is extended so an AssetBase extra (the stand) can be written with set_world_poses. The composition boundaries look right — using the existing layout_pose_to_scene_writes extension point is the intended pattern — but the stand-aware write dispatch was only added to one of the two reset paths.

Findings

🟡 Warning: embodiments/droid/droid.py:144 — layout_pose_to_scene_writes now emits a ("stand", …) write for every reset path, but only the relation-solver path (placement_events._write_scene_root_pose_to_sim) handles the stand's AssetBase via set_world_poses. The fixed-pose reset path (terms/events._write_scene_pose, built when a Droid is given an explicit initial_pose) still calls write_root_pose_to_sim unconditionally and will raise AttributeError on reset. The kitchen example is unaffected (it places via relations), but a fixed-pose Droid regresses. The two write routines are near-duplicates that have diverged — worth a single shared dispatch helper.

🔵 Improvement: tests/test_droid_stand_height.py:139 — the kitchen reset check only compares robot/stand z. Since the stand and robot are written through different Isaac APIs (set_world_poses vs write_root_pose_to_sim) and the spec applies a 1.57 rad yaw, a quaternion-order mismatch would leave the stand mis-rotated while z still matches. Consider also asserting the stand's world orientation equals the robot's.

Test Coverage

Good coverage overall: the consolidated test_droid_stand_height.py exercises stand-height scaling, robot/stand pose wiring, placement bbox, and an end-to-end kitchen reset, all under one run_simulation_app_function session (correct inner/outer pattern). test_placement_events.py adds a mock-env test that the Droid robot+stand are both written, plus a missing-asset assertion. Two gaps: (1) no test covers the fixed-pose (reset_placement_asset_pose) Droid path that the finding above breaks; (2) reset alignment only checks z, not orientation.

Verdict

Minor fixes needed — mainly routing the fixed-pose reset path through the same stand-aware write dispatch.