Review Summary

This is a well-targeted fix for #4970 — the stale sensor data issue after environment reset. The approach is minimal and correct.

The Bug (Recap)

When scene.reset(env_ids) runs inside ManagerBasedRLEnv._reset_idx:

1. Each sensor's reset() zeros its _data buffers ✓
2. reset_envs_kernel then marked the env as outdated → ❌ Problem
3. Next sensor.data access triggered a lazy refetch from a physics buffer that hadn't been stepped
4. The zeros got overwritten with stale pre-reset values

The Fix

# Before: is_outdated[env] = True  (triggering unwanted refetch)
# After:  is_outdated[env] = False (preserving the zeroed data)

This is the correct single-point fix. The flag is re-armed by InteractiveScene.update after the next physics step, restoring normal lazy-fetch behavior.

✅ Strengths

1. Minimal diff: One line change in reset_envs_kernel covers all affected sensors (ContactSensor, IMU, PVA, JointWrenchSensor) across both PhysX and Newton backends.

2. Comprehensive test coverage: 8 new regression tests (4 sensors × 2 backends) that specifically reproduce the manager-based RL flow. Each test was verified RED→GREEN.

3. Well-documented: The changelog fragment and updated docstring in reset_envs_kernel clearly explain the rationale.

4. No API changes: Purely internal behavior fix — no breaking changes for downstream users.

⚠️ Considerations

1. Sensors Without Zeroing reset() Override

The PR description notes:

Sensors that don't zero data (e.g. ray-caster, frame transformer) still see refetches at the next update-period boundary.

This is correct because those sensors don't override reset() to zero their data — they rely on the fetch. But it's worth confirming this behavior is still exercised by existing tests.

Question for author: Were the existing ray-caster and frame-transformer tests confirmed to still pass with this change? The PR description mentions existing ContactSensor (20) and IMU (8) tests passed, but doesn't explicitly mention ray-caster/frame-transformer.

2. Partial Reset Edge Case

With scene.reset(env_ids) where env_ids is a subset:

• Only the specified envs get is_outdated[env] = False
• Other envs retain their is_outdated state

This looks correct, but the new tests only cover single-env scenarios. It would strengthen confidence to see test_reset_with_env_ids_only_zeros_selected_envs patterns also check that non-reset envs retain their normal lazy-fetch behavior (i.e., their is_outdated flag isn't accidentally cleared).

3. Missing Import in Newton IMU Test

Looking at test_imu.py (Newton), the test uses RigidObject but I see it's correctly imported:

from isaaclab.assets import RigidObject, RigidObjectCfg

Good — this was updated.

🔍 Regression Test Robustness

The tests correctly mirror the _reset_idx flow:

1. Run simulation until sensor has non-zero data
2. Write new asset state
3. Call scene.reset(env_ids) without a subsequent physics step
4. Assert sensor.data returns zeros, not stale values

This is the exact scenario from issue #4970. The assertion style is clear:

assert post_reset_force_mag == 0.0, (
    "Contact sensor returned stale pre-reset data after scene.reset(): "
    f"got {post_reset_force_mag}, expected 0.0"
)

Verdict

Looks good to merge pending:

1. Confirmation that ray-caster/frame-transformer tests still pass (should be covered by CI)
2. (Optional) Consider adding a multi-env partial-reset test to verify non-reset envs retain outdated flag

The fix is correct, minimal, and well-tested. Nice work @AntoineRichard!

📋 Update (2026-05-27)

New commits pushed: 153a1e5c → ce490ebe

Reviewed the diff between previously-reviewed commit and new head. The changes appear to be unrelated to this PR's fix (stale sensor data after reset). They include:

• CI: Added ovphysx pip package to rendering workflow
• Docs: LEAPP export updates, Windows tab-sets, ecosystem link fixes
• New: PPISP camera QA demos (ppisp_camera.py, ppisp_camera_ovrtx.py)
• Cleanup: Removed legacy OVRTX 0.2.x code paths
• Fix: Tutorial CPU tensor device mismatch (add_new_robot.py)
• Test: Docstring updates, golden images for ovphysx rendering

These look like upstream base-branch changes that merged in, not modifications to the sensor reset fix itself. The core fix in reset_envs_kernel (flipping is_outdated[env] = False) remains unchanged.

Original review assessment stands: ✅ Ready to merge.