[hougantc-nvda]

Description

Review stack · 3/5: #8787 → #8788 → #8789 → #8790 → #8791

Review only this PR’s changes. The PRs target main; the comparison isolates this layer from its predecessors.

Enable the registered GR1T2, G1 locomanipulation and fixed-base G1 tasks to run with kitless Newton or CPU OvPhysX. Explicit physics selectors choose matching task presets and USD variants; omitting the selector preserves Kit defaults.

• Task ownership: robot mappings, action conditioning, model assets and backend tuning stay in isaaclab_tasks. Teleop consumes those configurations through the existing adapter.
• Joint ordering: preserve the configured Pink joint order so hand targets reach the intended joints.
• Shared assets: extract the steering-wheel helper used by GR1T2 and G1, retaining graspable rim/spokes and valid collision shapes without a cross-task import.

The existing G1 head camera remains available at this PR boundary; wrist cameras and presentation changes belong to #8790.

Registered humanoid tasks prepare their selected backend before scene construction through both gym.make() and instantiate(cfg). Repeated preparation preserves caller overrides and copied settings. Declared spawn settings and module-level callbacks preserve prepared-config serialization/restoration. GR1T2 uses normal independent configuration copies. Its enabled action reads conditioned hand limits and mimic metadata from the actual asset-owned spawn result, so restoring a copied configuration does not mutate its source. Changing backends after preparation requires a fresh configuration. Task construction tests remain independent of optional teleop imports; the XR integration harness stays in isaaclab_teleop.

See registered tasks and physics selectors.

Dependencies: public isaaccapture==1.7.93a1 from NVIDIA’s Python index, using the optional XR selection documented in #8788. The stack includes upstream's merged provider upgrade #8858; the installation recipe reuses its OVRTX, OVStage, OvPhysX and schema declarations. Project-wide Kit/provider defaults are unchanged.

Backend preparation preserves required physics/action setup when optional teleoperation configuration is absent. Input bounds are named with their units and purpose; their values and physics tuning are unchanged.

Type of change

• New feature
• Bug fix
• Documentation update

GR1T2 presets use the supported joint_effort_limit field with the task-specific tuned values.

Testing

At this PR boundary (f610df778807): 231 passed, 2 skipped, 10 deselected, 36 warnings in the selected renderer, adapter, live/replay and task-owner suites applicable to this layer. These boundary runs overlap; their counts are not additive. Repository formatting/lint and the Python 3.12 changelog check pass. The reused public-package environment passes uv pip check for all 268 installed packages.

The task regressions exercise real registered constructors, public config serialization/restoration, and the actual Asset copy/spawn boundary. The GR1T2 action consumes the spawned asset’s conditioned hand metadata without custom copy semantics. G1 bounds/validation tests cover multiple environments and non-identity joint ordering.

Cumulative runtime integration at 4e608f200e21: public Capture 1.7.93a1 and the published CLI passed IWER scene/control checks for GR1T2 CPU OvPhysX, GR1T2 split-GPU Newton, fixed-base G1 split-GPU Newton, G1 locomanipulation CPU OvPhysX with --headless, and SO101 CPU OvPhysX. Both reconstructed eyes, meaningful motion and Reset were inspected. Requested PiP feeds and the default Vulkan mirror worked; closing the mirror shut down each non-headless instance normally. Headless G1 created no mirror even with the explicit mirror flag, retained PiP/streaming and exited normally after a deliberate owner interrupt. GR1T2 Newton and headless G1 also passed disconnect/reconnect and resumed motion. SO101 passed clutch/gripper checks; cube retention is not qualified.

With the same packages, hardware, input, seed and warm caches, GR1T2 CPU delivered 49.2 median decoded video frames/sec versus baseline runs of 49.0 and 50.4; fixed-base G1 Newton delivered 32.8 versus 33.0. Each median uses three 10-second steady windows. These bounded comparisons showed no material streaming slowdown; they do not measure headset FPS or establish a full physics-performance guarantee.

Fresh physical-headset, full native Kit/Isaac Sim, and native MCAP-file replay qualification were not performed in this follow-up. The historical cold-attach and intermittent GR1T2 first-Play reports remain open; successful warm runs do not establish a fix. Existing native USD/asset and split-GPU camera-device warnings remain visible.

Checklist

• I have read and understood the contribution guidelines
• I have run the repository formatting checks and the separate Python 3.12 changelog check
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have validated the changed behavior with focused coverage and applied the test-audit criteria to any test changes
• I have added a changelog fragment under source/<pkg>/changelog.d/ for every touched package (do not edit CHANGELOG.rst or bump package versions — CI handles that)
• I have added my name to the CONTRIBUTORS.md or my name already exists there

The warnings checkbox remains open: native-runtime warnings and the separate Kit dependency selection are not fully qualified by this run.

[hougantc-nvda]

@greptile Please review the latest changes in this PR.

[greptile-apps]

Confidence Score: 4/5

[High impact] The PR reaches registered tasks, shared rendering, simulation setup, and live and replay teleoperation.

Update the failed-seal test renderer before merging; both cases now stop before reaching their checks.

Findings

1.  Publication tests stop early ▶

Summary

Enables G1 and GR1T2 tasks with kitless physics while retaining Kit defaults.

• G1 and GR1T2 tasks use selected physics backends with tuned robot controls.
• An external XR session can borrow Isaac Lab’s renderer and scene through a lease.
• Live input and MCAP replay can use OVXR with the simulation’s scene.
• Kitless runs and headless Kit XR keep their runtime ownership with the selected application.
• The environment browser keeps its existing behavior as its page updates.

The supplied previous threads are unnumbered. The scene-check, live-reset and standalone-startup fixes are present. hougantc-nvda disputed the replay-reset finding because the supported provider admits forced requests synchronously unless shutdown is latched; no new evidence contradicts that explanation.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Resolved task configuration] --> B[HumanoidTeleopEnv]
    B --> C{Selected physics}
    C -->|Kit| F[Base environment construction]
    C -->|OvPhysX or Newton| D[Prepare task backend]
    D --> E[Check prior preparation and bind spawn callback]
    E --> F
    F --> G[Create scene and action managers]

Loading

_{Reviews (4) · Last reviewed commit: "Prepare registered humanoid tasks before..." · Reviewed by Greptile}

[hougantc-nvda]

@greptile Please review the current head (b1de1f7) again. Recheck the fixes and remaining issues against the published code, including the actual interfaces and callers at this PR boundary.

[hougantc-nvda]

@greptile please review this PR again at the latest head (d337b9b0cb91106142de0cbdd01256058f33e10f), including the changes since your previous review.

[isaaclab-review-bot]

Isaac Lab Review Bot

All reviewed code categories look good.

• PR description: The title and body understate the renderer, replay, SO101, launcher, and visualizer changes; summarize those areas and report validation performed on the final head.
• Design and architecture: Architecture looks good.
• Scope and over-engineering: No over-engineering issues.
• API: API changes look good.
• Compatibility and deprecation: No breaking changes identified.
• Implementation: Implementation looks good.
• Style consistency: Style consistency looks good.
• Test quality: Test quality looks good.

No blocking issues. No inline findings.

The full PR diff was reviewed; some supplemental surrounding file context was omitted.

Automated review; human maintainers own approval decisions.

[hougantc-nvda]

@isaaclab-review-bot review