[Tasks] Fix surface gripper observation shapes

[StafaH]

Description

Fix surface-gripper stack observations so object_grasped and object_stacked return one boolean per environment. The surface-gripper state was reshaped to (num_envs, 1) before being combined with an (num_envs,) predicate, which caused PyTorch to broadcast the result to (num_envs, num_envs) while stepping the environment.

No new dependencies are required.

Reported in TC_146506 for IsaacContrib-Stack-Cube-UR10-Long-Suction-IK-Rel with the Newton Kamino backend.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Release backport

• Backport this pull request to the active release branch after it merges into develop

Screenshots

Not applicable; this corrects an observation tensor shape.

Validation

• uv run python tools/changelog/cli.py check --include-worktree develop
• uv run isaaclab -f

Checklist

• I have read and understood the contribution guidelines
• I have run the pre-commit checks with uv run isaaclab -f
• Documentation changes are not required for this internal shape correction
• My changes generate no new warnings
• I have added a changelog fragment under source/isaaclab_tasks/changelog.d/
• My name already exists in CONTRIBUTORS.md

[isaaclab-review-bot]

Isaac Lab Review Bot

The patch corrects unintended cross-environment broadcasting in the surface-gripper branches of object_grasped and object_stacked by flattening gripper state to one value per environment. The focused regression test covers both predicates across three distinct environment states, and the user-visible fix has an appropriate changelog fragment.

• Design and architecture: The change preserves the existing observation and scene-access structure while aligning surface-gripper predicates with the flat per-environment tensors already produced by the geometry and joint-gripper paths. No new abstraction or dependency is introduced.
• API: Function signatures, configuration inputs, and boolean result dtype remain unchanged. The corrected return shape is (num_envs,) rather than the accidentally broadcast (num_envs, num_envs), restoring the intended per-environment observation behavior.
• Implementation: The changed reshape directly prevents PyTorch from broadcasting (num_envs, 1) against (num_envs,). The test independently distinguishes gripper states and object geometry, validating expected values as well as shape for both predicates. Its mocked Warp conversion is appropriately focused on tensor-combination semantics, though integration with the real Warp state layout remains outside this test's scope.

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.

[greptile-apps]

Greptile Summary

This PR fixes unintended cross-environment broadcasting in two surface-gripper stack observations.

• Flattens surface-gripper state to one scalar per environment before combining predicates.
• Adds a three-environment regression test with distinct gripper states and object geometry.
• Adds a changelog fragment describing the corrected observation shape.

Confidence Score: 5/5

The PR appears safe to merge with no actionable correctness, security, or repository-rule issues identified.

The new one-dimensional reshape matches the surface-gripper state contract and prevents the documented quadratic broadcast, while the regression test covers both affected predicates.

Important Files Changed

Filename	Overview
source/isaaclab_tasks/isaaclab_tasks/contrib/stack/mdp/observations.py	Corrects surface-gripper state reshaping in object_grasped and object_stacked to preserve the environment batch.
source/isaaclab_tasks/test/contrib/stack/test_observations.py	Adds a simulation-independent regression test verifying per-environment shapes and predicate values.

_{Reviews (1): Last reviewed commit: "Fix surface gripper observation shapes" | Re-trigger Greptile}

[StafaH]

run-ci