Add G1 AGILE tabletop apple-to-plate environment

[lgulich]

Problem

IsaacLab-Arena needs a tabletop manipulation task where the G1 robot uses the WBC-AGILE locomotion policy to pick up an apple and place it on a plate, while balancing in place.

Ref: ISAAC-12630

Solution

Add a new G1AgileTabletopAppleToPlateEnvironment that wires the G1WBCAgileJointEmbodiment (from PR #489) with the existing PickAndPlaceTask, a Seattle Lab table scene, and appropriate object assets.

Changes

• isaaclab_arena_environments/g1_agile_tabletop_apple_to_plate_environment.py — New environment class: G1 robot at (-0.4, 0, 0) facing a table with an apple (pick object) and a clay plate (target). Uses G1WBCAgileJointEmbodiment for balance + upper body control. 30-second episodes. Supports --object, --embodiment, --teleop_device CLI args.
• isaaclab_arena_environments/cli.py — Register the new environment in the ExampleEnvironments dict.
• isaaclab_arena/tests/test_g1_agile_tabletop_apple_to_plate.py — Two tests: (1) initial state is not terminated (apple starts away from plate), (2) teleporting apple onto plate triggers success termination. Uses correct base-height command (0.75) to keep the robot stable.

Testing

• New unit tests added (2 tests)
• Linters pass locally (black, flake8, isort, pyupgrade, codespell, license headers)
• CI pipeline (tests require Isaac Sim Docker with GPU)

Notes

• Object positions (apple, plate, robot) are based on Seattle Lab table geometry and G1 arm reach. May need visual tuning in simulator.
• No new task class needed — the existing PickAndPlaceTask handles contact-sensor success detection, object-dropped termination, and metrics.
• Self-review caught and fixed a test issue: the initial-state test was sending zero base-height commands, which would cause the robot to squat. Fixed to use 0.75 (matching established pattern from test_g1_wbc_embodiment.py).

---

Generated by autodev — Claude Code

[isaaclab-review-bot]

Review Summary

Clean new environment addition that follows established patterns well. The G1AgileTabletopAppleToPlateEnvironment correctly wires G1WBCAgileJointEmbodiment with PickAndPlaceTask and the CLI/test integration is solid. Pre-commit and docs CI pass. Two minor suggestions below.

✅ What looks good

• Pattern adherence: Environment class, CLI registration, and test structure all follow the established conventions (matches KitchenPickAndPlaceEnvironment, GalileoPickAndPlaceEnvironment, etc.)
• Correct base-height command (0.75): Both tests properly set actions[:, -4] = 0.75 to keep the AGILE WBC robot standing, matching the pattern established in test_g1_wbc_embodiment.py
• Good test design: The two-test approach (initial state not terminated + teleport-to-success) provides meaningful coverage of the task lifecycle without requiring full manipulation
• Clean reuse: No new task class needed — PickAndPlaceTask handles everything

📝 Suggestions

1. Task description is hardcoded to "apple" but --object is configurable

File: g1_agile_tabletop_apple_to_plate_environment.py, line 66

The task_description always says "Pick up the apple from the table and place it onto the plate" even when a different object is passed via --object. Consider making it dynamic:

task_description=f"Pick up the {args_cli.object.replace('_', ' ')} from the table and place it onto the plate.",

Or if the description is only used for display/logging and the default object is always apple, this is fine as-is — just worth noting.

2. Test duplicates environment setup instead of reusing the environment class

File: test_g1_agile_tabletop_apple_to_plate.py

The test's get_test_environment() manually reconstructs the entire environment (assets, poses, task, embodiment) rather than instantiating G1AgileTabletopAppleToPlateEnvironment and calling get_env(). This means if positions or assets change in the environment class, the test won't reflect those changes.

That said, I see other tests in this repo follow the same pattern (e.g., test_g1_wbc_embodiment.py), so this may be an intentional convention to keep test setup explicit. If so, no change needed — just flagging the drift risk.

Verdict

LGTM with minor nits. This is a straightforward, well-structured environment addition. The two suggestions above are non-blocking. 👍

[isaaclab-review-bot]

🤖 Isaac Lab Review Bot

Reviewed 5 changed files. One bug and a couple of style nits:

🐛 Bug: Missing @register_asset decorator

isaaclab_arena/embodiments/g1/g1.py — G1WBCAgilePinkEmbodiment is missing the @register_asset decorator. Every other embodiment class (G1WBCJointEmbodiment, G1WBCPinkEmbodiment, G1WBCAgileJointEmbodiment) has it. Without it, asset_registry.get_asset_by_name("g1_wbc_agile_pink") will fail at runtime.

# Missing this:
@register_asset
class G1WBCAgilePinkEmbodiment(G1EmbodimentBase):

📝 Style nits (non-blocking)

1. g1_agile_tabletop_apple_to_plate_environment.py — import isaaclab.sim as sim_utils and from isaaclab_arena.utils.pose import Pose are placed mid-function between asset creation statements (lines ~36 and ~51). Move them to the top of get_env() with the other imports for readability.

2. Retargeter — G1WbcAgilePinkIsaacTeleopRetargeter imports the same _build_g1_pink_locomanipulation_pipeline as the non-AGILE retargeter. This is fine if the pipeline is truly the same for both, but worth a comment noting that the AGILE variant shares the same teleop pipeline.

Overall the environment follows existing patterns well. Tests are solid — the warmup + re-place strategy for the initial state test is a nice touch. Just needs the decorator fix.

[lgulich]

Fix: False success termination in pick-and-place tasks

Root Cause

The success termination was firing incorrectly — PhysX GPU pipeline reports phantom contact-sensor forces between objects that are far apart. The contact sensor reported 300-3700 N between the apple and plate even when they were 0.3-1.7m apart. When the apple had low velocity (e.g., at the peak of its trajectory), both conditions for success were satisfied, triggering false success termination.

Fix

Added a proximity guard to object_on_destination termination:

• New optional parameters: destination_cfg (SceneEntityCfg) and max_distance (float)
• When provided, verifies the pick-up object is within max_distance of the destination before contact force is considered valid
• Backward compatible — existing callers see no behavior change
• Enabled in PickAndPlaceTask with max_distance=0.15m

Test Results

• test_initial_state_not_terminated: PASSED ✅ (was failing)
• test_apple_on_plate_succeeds: PASSED ✅ (still passes)
• test_agile_standing: PASSED ✅ (no regression)

Video

Environment recording attached — shows G1 robot standing at the table with the apple, plate, and table visible. AGILE WBC policy maintains balance.

---

Generated by autodev — Claude Code

[greptile-apps]

Greptile Summary

This PR adds a new G1AgileTabletopAppleToPlateEnvironment that wires G1WBCAgileJointEmbodiment with the existing PickAndPlaceTask, a maple_table_robolab scene, and an apple-to-plate manipulation task. It also adds G1WBCAgilePinkEmbodiment, an OpenXR retargeter for the AGILE+Pink configuration, a proximity guard to object_on_destination to suppress spurious contact forces from the GPU physics pipeline, and two simulation tests for the new environment.

Confidence Score: 5/5

Safe to merge; all findings are P2 style/quality suggestions that do not block correct runtime behaviour.

No P0 or P1 defects were found. The proximity guard fix is backward-compatible and the new environment follows established patterns. The three P2 comments (apple scale, hardcoded re-placement position, simplified test scene) are informational quality improvements that should not block the PR.

isaaclab_arena_environments/g1_agile_tabletop_apple_to_plate_environment.py — verify the apple scale=(0.01,0.01,0.01) is correct for the USD asset's native units.

Important Files Changed

Filename	Overview
isaaclab_arena_environments/g1_agile_tabletop_apple_to_plate_environment.py	New environment class; apple is instantiated with scale=(0.01, 0.01, 0.01) — 100× smaller than the default — which may not match the actual USD asset units.
isaaclab_arena/tests/test_g1_agile_tabletop_apple_to_plate.py	Two simulation tests added; helper hardcodes apple's initial position rather than deriving it from the asset, and the scene layout differs from the production environment.
isaaclab_arena/tasks/terminations.py	Proximity guard added to object_on_destination; handles XformPrimView destinations defensively; logic is correct for the new environment's RigidObject destination.
isaaclab_arena/tasks/pick_and_place_task.py	Proximity guard is now opt-in via success_proximity_max_distance; existing callers with default 0.0 are unaffected; clean backward-compatible change.
isaaclab_arena/embodiments/g1/g1.py	G1WBCAgilePinkEmbodiment added alongside the existing G1WBCAgileJointEmbodiment; both follow established patterns.
isaaclab_arena/assets/retargeter_library.py	New G1WbcAgilePinkIsaacTeleopRetargeter registered for the g1_wbc_agile_pink embodiment; follows existing retargeter pattern.
isaaclab_arena_environments/cli.py	G1AgileTabletopAppleToPlateEnvironment correctly registered in ExampleEnvironments dict.

Sequence Diagram

sequenceDiagram
    participant CLI as CLI / ArenaEnvBuilder
    participant Env as G1AgileTabletopAppleToPlateEnvironment
    participant Scene as Scene (maple_table_robolab + apple + plate)
    participant Task as PickAndPlaceTask
    participant Term as object_on_destination
    participant Prox as proximity guard

    CLI->>Env: get_env(args_cli)
    Env->>Scene: assemble assets + On(table_reference) relations
    Env->>Task: PickAndPlaceTask(success_proximity_max_distance=0.15)
    Task->>Term: register termination with destination_cfg + max_distance
    CLI->>Scene: ArenaEnvBuilder.make_registered()
    loop Each physics step
        Term->>Term: check contact force > 1 N
        Term->>Term: check object velocity < 0.1 m/s
        Term->>Prox: distance(apple, plate) < 0.15 m ?
        Prox-->>Term: close_enough
        Term-->>CLI: condition_met (success)
    end

Loading

_{Reviews (1): Last reviewed commit: "fix(test): enable proximity guard in tab..." | Re-trigger Greptile}

[greptile-apps]

Verify apple scale is intentional

scale=(0.01, 0.01, 0.01) reduces the asset to 1% of its original size. No other environment in the codebase applies a scale this small to a pick-up object — the analogous PickAndPlaceMapleTableEnvironment uses the default scale of 1.0. If the apple_01_objaverse_robolab USD is already in meter units an apple would end up roughly 1 cm in diameter, which would make grasping and contact detection unreliable. If the asset is authored in centimetres then this conversion is correct; please add a brief comment explaining the intent so future readers don't wonder whether this is a typo.

[greptile-apps]

Hardcoded re-placement position may silently drift from the test setup

init_pos is hardcoded to [0.15, 0.15, 0.05] instead of being derived from the apple asset configured in get_test_environment. If the initial pose on line 37 of that function is updated, these two values will silently diverge and the "re-placed" apple will be at the wrong location, potentially making the test pass or fail for the wrong reason. Consider reading the value from the apple object or from a shared constant:

# option: derive from apple's stored initial pose
init_pos = torch.tensor([apple.initial_pose.position_xyz], device=env.unwrapped.device)

[greptile-apps]

Test scene layout differs significantly from the production environment

The test uses the generic table background with the robot at (-0.4, 0, 0) facing +x (identity quaternion), whereas the production environment uses maple_table_robolab with the robot at (1.2, 0, 0) facing -x (180° z-rotation). The tests exercise the underlying component logic well, but they will not catch issues specific to the Seattle Lab table geometry, the actual G1 arm-reach envelope from that position, or the On(table_reference) spatial relation used in the environment. A short comment here explaining why a simplified layout is used would help future contributors understand the coverage gap.

[lgulich]

Hi @alexmillane — friendly reminder that this PR is ready for your review. All CI checks are passing (pre-commit, tests, GR00T policy tests, docs build). Let me know if you have any questions!

---

Generated by autodev — Claude Code