Superseded by focused CI fixes

[ooctipus]

Superseded by independently reviewable fixes:

• Fix operational-space feedback and task-frame consistency #7868: operational-space controller test feedback and convergence.
• Let standalone demos complete the health check after startup #7869: standalone demos finishing their health check after the startup deadline.
• Fix Franka Lift reset sampling after the collider default changed #7870: Franka Lift/Reorient collider selection and reset sampling.
• Fix legacy OVRTX cloning of instanced geometry and unused variants #7871: legacy OVRTX instance cloning, including the inactive-variant regression discovered in this PR.
• Fix Franka Pour collider variant #7866: the independently authored Franka Pour collider fix.

This combined PR is closed in favor of those focused changes. Each description records the confirmed introducing commit or external asset change, affected CI failures, validation, and remaining uncertainty.

CI on this branch passed the controller, RL export, Pour, and standalone jobs. Its renderer change fixed the 12 Franka image comparisons but introduced four heterogeneous Kuka failures by visiting intentionally unused clone-source paths; #7871 corrects that renderer-boundary mistake and validates all 16 cases against unchanged reference images.

Two additional failures investigated alongside these fixes have separate PRs: #7872 selects the CUDA runtime matching PyTorch, fixing Newton graph capture with the CUDA 13 wheels used by #7674; MuJoCo-Warp #1675 removes duplicate CCD compilation during cold startup. The latter reduces measured startup time locally, but its effect on Isaac Lab's cloud startup timeouts still needs CI confirmation.

[greptile-apps]

Confidence Score: 4/5

The PR appears safe to merge, with a non-blocking hardening issue around non-finite standalone soak configuration.

Findings

1.  Non-finite soak can hang ▶

Summary

This PR addresses shared Franka asset, legacy rendering, operational-space test, and standalone smoke-supervisor regressions.

• Selects convex-hull colliders for Franka Pour, Lift, and Reorient configurations.
• Expands nested USD instances only in the legacy OVRTX export session.
• Aligns operational-space feedback with link-origin conventions and stabilizes nullspace convergence.
• Allows a ready standalone process to complete its configured soak interval after the startup deadline.
• Adds focused regression coverage for the changed configurations and execution paths.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Launch standalone process] --> B{Process ready?}
  B -- No --> C{Startup timeout reached?}
  C -- Yes --> D[Terminate process group]
  C -- No --> B
  B -- Yes --> E[Begin soak interval]
  E --> F{Soak interval elapsed?}
  F -- No --> E
  F -- Yes --> D

Loading

_{Reviews (1) · Last reviewed commit: "Allow standalone demos to finish soaking..."}

[greptile-apps]

Non-finite soak can hang

Once readiness is observed, the startup deadline is disabled and soak_time becomes the only remaining bound. Because ISAACLAB_STANDALONE_SOAK_TIME is parsed with float(...) without a finiteness check, values such as inf or nan make the soak comparison permanently false. A healthy but still-running demo would then remain supervised indefinitely, so the configured soak interval should be validated as finite.

[isaaclab-review-bot]

Isaac Lab Review Bot

Reviewed the four targeted regression fixes: task-local Franka collider selection, legacy OVRTX export-session uninstancing, operational-space test feedback corrections, and post-readiness smoke-test soaking. The changed paths are scoped to their stated consumers and include focused regression coverage.

• Design and architecture: Nested USD instances are expanded only on the anonymous legacy OVRTX export stage when keep_env_roots is enabled, preserving instancing on the simulation stage, single-environment exports, and the ovstage path. Franka collider selection remains localized to Pour and rigid Lift/Reorient configurations rather than changing the shared asset default.
• API: No public interface is removed or renamed. export_stage_to_string retains its signature and default behavior boundary, while the smoke supervisor change refines the existing startup-timeout semantics so readiness transitions control to the soak interval. Package-specific changelog fragments are present.
• Implementation: The OVRTX regression test verifies exported geometry, path-specific instance preservation, and unchanged input layers. Collider tests verify the selected USD variant for Pour, Lift, and Reorient, while the supervisor test covers readiness immediately before the startup deadline. Operational-space feedback now consistently uses link-origin velocity data alongside the link-origin Jacobian and pose. A non-blocking tradeoff is that each task assigns the complete spawn.variants mapping, so future presets that replace that mapping wholesale must preserve the collider selection.

No blocking issues. No inline issue met the actionable-evidence threshold; the assessment above records the review feedback.

Automated review; human maintainers own approval decisions.

[maxkra15]

did you check that this doesnt create regressions in other tasks?

[maxkra15]

thanks, was working on this as well

[kellyguo11]

run-ci