Batch Support for AxisAlignedBoundingBox

[zhx06]

Summary

Add batch dimension support to AxisAlignedBoundingBox

Detailed description

• What was the reason for the change?
  • To support multi-environment batched placement, AxisAlignedBoundingBox needs to handle N bounding boxes simultaneously instead of only one.
• What has been changed?
  • Refactored AxisAlignedBoundingBox from a dataclass to a class with (N, 3) float32 tensor storage. Returns tuples/floats for N=1, tensors for N>1.
  • Fixed float32 rounding issue in _validate_on_relations Z boundary check.
  • Added test_bounding_box.py for single-env and multi-env coverage.
• What is the impact of this change?
  • No behavior change for existing code: backward compatible for single env (N=1).
  • Enables future multi-env batch support for the coming relation solver MR.

[cvolkcvolk]

I feel like the dual-mode return type is a reliability issue. Properties return tuple/float when N=1 and tensor when N>1. Code that works in single-env silently changes behavior when N increases. Recommend to always returning tensors

[cvolkcvolk]

Additionally we could simplify below some parts like self._min_point = self._to_batched_tensor(min_point),as its always a tensor (the input as well).

[cvolkcvolk]

This would require some updates downstream, but it would be worth it in my opinion

[cvolkcvolk]

Suggestion to rename this function to what is formatted

[cvolkcvolk]

However can remove this if the class only handles tensors

[cvolkcvolk]

I think this would break silently for N>1? When N>1, min_point returns a tensor and the subsequent torch.rand(3) arithmetic produces wrong-shaped results ((N,3) instead of (3,)).
I would suggest to move to single input and output type (tensors)

[kellyguo11]

🤖 Isaac Lab Review Bot

Summary

This PR refactors AxisAlignedBoundingBox from a @dataclass with tuple storage to a class with (N, 3) float32 tensor storage, enabling future multi-env batched placement. All 14 changed files are consistent updates propagating the new tensor API. CI all green.

Design Assessment

Design is sound. The always-tensor approach is clean — no dual code paths, no conditional returns based on N. The _to_batched_tensor helper keeps the constructor ergonomic for the N=1 case. Broadcasting in overlaps() is correct for N-vs-1 comparisons.

Findings

🟡 Warning: _to_batched_tensor doesn't handle list input — bounding_box.py:49-54
While the type hints specify tuple | Tensor, passing a Python list would raise AttributeError: 'list' object has no attribute 'dim'. Consider accepting Sequence[float] or adding a list→tuple/tensor conversion for robustness:

if isinstance(value, (tuple, list)):
    return torch.tensor([value], dtype=torch.float32)

🟡 Warning: _validate_on_relations Z-fix changes semantics subtly — object_placer.py:280-286
The old code used translated() world bboxes (float32 tensor math), the new code manually adds Python floats. This fixes float32 rounding ✅ but the X/Y checks still use the translated tensor path (child_world.min_point[0, 0]). Consider making X/Y consistent with Z for uniform precision, or add a comment explaining why Z needs special treatment.

🔵 Suggestion: Constructor could validate min ≤ max — bounding_box.py:37-38
The constructor asserts shape correctness but not value correctness. min_point > max_point per-axis creates an inverted bbox that overlaps(), size, etc. interpret incorrectly. Pre-existing issue, but with tensor storage a simple assert (self._min_point <= self._max_point).all() would catch bugs early.

🔵 Suggestion: get_random_pose_within_bounding_box only uses env 0 — bounding_box.py:256
This function takes bbox.min_point[0] / bbox.max_point[0], which is correct for N=1 but silently ignores other envs for N>1. Consider adding assert bbox.num_envs == 1 or an env_idx parameter.

Test Coverage

• New tests: test_bounding_box.py — 8 tests covering single-env and multi-env properties, transforms, overlaps, and corners. Good coverage.
• Updated tests: test_no_collision_loss.py and test_object_placer_init.py properly updated for [0, axis] indexing.
• Minor gap: No test for scaled() with batched (N>1) scale tensor, and no regression test for the float32 Z-rounding fix. Not blocking.
• Verdict: Coverage adequate. Test quality is good.

CI Status

All checks passing ✅

Verdict

COMMENT

Clean, well-executed refactor. The tensor-first approach is the right call for future batch support. No critical issues — the warnings are minor and the suggestions are nice-to-haves.

---

Review by Isaac Lab Review Bot — [Expert Analysis] + [Error Handling Audit] + [Test Coverage Check]