[yuecideng]

Description

A camera whose view direction is collinear with its up vector currently switches the entire look_at_to_pose batch to a fallback direction. Changing only camera B can therefore rotate unchanged camera A by 90 degrees, or introduce NaNs into another camera facing along the fallback axis.

Apply the fallback only to the affected rows and choose the less-aligned X/Y reference axis for each row. Preserve the existing default-up vertical-camera orientation. Extend the existing math test to compare mixed batches against isolated calls, cover custom per-camera up vectors on CPU/CUDA, and check known camera orientations.

Dependencies: none.

Type of change

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

Validation

• Regression check before the production fix: all four CPU/CUDA/default-up/custom-up cases failed.
• python -m pytest -q tests/utils/test_math.py tests/sim/sensors/test_camera_reset.py tests/gym/envs/managers/test_randomize_camera_extrinsics.py --run-gpu: 76 passed.
• black . and black --check --diff ., Black 26.3.1: passed.
• python docs/scripts/check_api_docs.py: 2805/2805 exports documented.
• git diff --check: passed.

The added coverage extends the existing camera math test, closing the missing batch-independence and custom-up degeneracy checks without introducing a separate fixture or test module.

Rendered comparison

Checked actual DexSim Hybrid offscreen camera frames with two arenas, CUDA look-at inputs, 480×480 RGB images, disabled denoising, a fixed scene, and zero physics steps. Only camera B's eye changes from oblique to vertical; camera A's inputs remain fixed.

Camera A response	Before	After
Rotation difference	90°	0°
Changed pixels	58.29%	0%
Mean absolute RGB difference (0–255)	36.9705	0

Documentation and context

The function docstring now states the per-camera fallback contract. No public exports or signatures change. The context impact command reports no affected topics; the sensor overview was also reviewed and its ownership/lifecycle guidance remains accurate, so no context edit is needed.

Checklist

• I have run the black . command to format the code base.
• I reviewed affected documentation and agent context, updated it where needed, or explained why no update was needed.
• Public API changes are reflected in the API docs (python docs/scripts/check_api_docs.py), if applicable.
• I verified appropriate behavioral coverage, reviewed the necessity of added or extended tests, and explained when no new tests were needed.
• Dependencies have been updated, if applicable (no dependency changes).

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge.

Summary

look_at_to_pose now applies its fallback only to cameras whose view and up directions are collinear.

• Look-at poses keep fallback choices within each camera.

No actionable issues found. This review checked the code and test expectations; it did not run tests.

_{Reviews (1) · Last reviewed commit: "fix(sensor): isolate batched look-at fal..." · Reviewed by Greptile}