[Camera] Remove remaining hot-path host syncs and overhead [7B/7] by StafaH · Pull Request #8105 · isaac-sim/IsaacLab

StafaH · 2026-09-28T05:42:55Z

Description

Part 7B of the camera performance series, based on the latest develop (including #8039, #8042, and #8087). A sweep of the per-step hot path found remaining host-device synchronizations and code that costs more than it gives. This PR fixes the verified findings: it removes the synchronizations, launch overhead, and dead or over-defensive code, and fixes the verified bugs.

Profiling on develop motivated this. The GPU sat idle 20–27% of each step, because the CPU blocked on synchronizations and then launched hundreds of small kernels. On Isaac-Velocity-Rough-AnymalD (4096 environments), for example, the CPU was blocked 12.8 ms of a 25.4 ms step.

Changes

Host synchronizations removed

Episode logging: the command, termination, and curriculum managers, survival_success_rate, the pose commands, and task success rates (Cabinet, Locomotion, Lift deformable/cable, Fourbar) log 0-d device tensors instead of reading back .item()/.cpu() on every reset. RewardManager already logged tensors, and the RL wrappers accept them. Curriculum states are logged as copies, because rsl_rl keeps episode infos for a whole rollout. The inspection accessor get_active_iterable_terms() still returns Python scalars.
Velocity command: heading control is branchless (torch.where) instead of is_heading_env.nonzero() every step. Output is identical.
Event ranges: core and Lift event terms use sample_uniform_from_ranges() to draw named components directly. One private, bounded bounds cache replaces both sets of tensor-building helpers; it is keyed by values, resolved device, and dtype. Curriculum edits take effect on the next call, and every call draws fresh samples in the existing order.
Ray caster: reset() no longer uploads its range tensor or samples drift when the drift ranges are zero (the default). In rough locomotion this reset runs almost every step. When drift is disabled after sampling, reset clears the selected environments’ old drift.
DiffIK: the if ee_quat.norm() != 0 guard in apply_actions, which synchronized every physics substep, is now a torch.where with the same all-or-nothing behavior.
Direct tasks: Reorient (about 14 host-list indexes per substep), Cabinet, and Locomotion store joint and body indices as device tensors, precompute static joint limits and default positions, and use the non-deprecated actuators.target_command API.
Other task terms:
- Locomotion walk_target_w caches its device target (it synchronized 4 times per step).
- Reorient's EpisodeErrorRecorder uses torch.where with NaN reductions instead of boolean masks.
- Lift reset events share the named-range sampler with the core event terms.
Joint limits: Newton, PhysX, and OvPhysX position-limit writes reuse the clamping counter and read it back only when the resulting message would be logged.

Launch overhead

RewardManager: each term writes its weighted value into a step-reward matrix column, followed by one sum and one batched episode-sum update. Reset is one mean and one fill. With 20 terms at 4096 environments, compute drops from 121 to 23 kernels (471 → 152 µs) and reset from 80 to 4 kernels (504 → 49 µs).
ObservationManager: the built-in noise functions (constant_noise, gaussian_noise, uniform_noise) are out-of-place, so the pre-noise clone and the second clone are skipped. Safe built-in noise functions are identified during _prepare_terms; custom noise functions, noise models, and modifiers still clone, including when a caller replaces a noise function after setup. The noise sampling order is unchanged.
CircularBuffer: one staged shift replaces K−1 slot copies for frames up to 2 MiB. Image-sized frames keep the slot loop, which is faster for them.
Lift contacts: one stacked norm replaces per-sensor norms. contacts() runs 5 times per step in Kuka Allegro, and this removes about 80 kernels per step.
ANYmal symmetry augmentation: one cached permutation and sign tensor per transform, bitwise-equal to the original, is 1.56 → 0.15 ms per call at 4096 environments.
RTX renderer: normals, motion vectors, rgb_hdr, and simple shading no longer copy each frame into a contiguous channel slice before reshaping.
Smaller changes:
- Reorient reads its command slices directly instead of concatenating them.
- sb3 computes dones on the host side.

Overengineered code removed

NewtonManager.step() ran the solver-internals reset every physics step. forward() already performs it when kinematics are dirty, so the extra call only launched with an all-False mask.
WrenchComposer.reset() returns early when inactive. Its buffers are already zero, and this removes about 7 memsets per rigid object per physics step (21 µs → 0.07 µs per call).
SensorBase: sensors with debug_vis=True no longer refresh on every physics substep. That refresh existed because debug callbacks once read private data. The callbacks now refresh the outdated buffers themselves. The original eager update was introduced in commit 4565902f5 to make visualization work without a data read; it did not pump rendering.
VisualizationMarkers.visualize() returns early when no backend is active.
Dead code: removed the Newton Warp renderer's output copy loop (every output is a zero-copy alias), the impossible PhysX ray-caster guard, and the redundant Lift body_state_b index conversion (redundant since Cache scene selections with automatic Torch dispatch #8087).

Bug fixes

Newton body_com_acc_w: the finite difference divided by the physics timestep even when an update spans several physics steps (Newton owning the decimation loop). That made the acceleration, and body_lin_acc_l2, decimation× too large. It now uses the elapsed time, matching joint_acc. The OvPhysX rigid-object implementation uses the same elapsed-time formula. PhysX and OvPhysX articulations read acceleration from the backend, so they do not need this finite-difference change.
reset_root_state_uniform built its range tensors in __init__ and ignored call-time pose_range/velocity_range, so curriculum updates were silently dropped.

Investigated and not included

Step reset, command resample, and interval-event nonzero: these are now the only synchronizations left in Isaac-Velocity-Rough-AnymalD. Each converts a mask into variable-length environment IDs consumed across the reset, command, and event APIs. Removing them requires a mask-based API redesign and is left for a separate design.
Needs a Kit trace before changing: the PhysX IMU/PVA per-update COM upload, PhysX contact-sensor per-substep fetches, and the Fabric render synchronizations. They rely on PhysX/Fabric semantics we have not measured. A similar "pageable upload synchronizes" assumption proved false for the Newton sensor flags.
Deferring the MPM sparse-grid status check: a solver reset clears the status, so a deferred check could lose capacity errors, and no test covers this path.
RMPflow guard: it already runs on the host, so removing the guard gains nothing. The joint_actions allocation is needed for LEAPP export tracing.
Lower confidence, left out: collection mask-write kernels and batching model-change notifications.

Type of change

Bug fix (non-breaking change which fixes an issue)
Behavior changes with migration notes in the changelog fragments:
- Episode log values are 0-d tensors; use float() or .item() where a Python number is needed.
- Reorient direct index attributes (actuated_dof_indices, finger_bodies, finger_wrench_bodies) and resolve_actuated_tendons ids are tensors.
- EpisodeErrorRecorder reports NaN for an empty selection after its first update.
- Custom sensors should refresh their buffers inside _debug_vis_callback.

Performance

RTX PRO 6000 Blackwell Workstation Edition (driver 595.91.07); PyTorch 2.12.0+cu130; Warp 1.17.0; Newton 1.6.0; GPU otherwise idle. develop at 60bc09f17 compared with this PR on that base, with the same uv.lock. The branch has since merged develop at f25d3f09b (#8042, #8090, #8102). The numbers below were measured before that merge; #8042 changes the Kuka camera observations and #8090 touches the simulation context, so they have not been re-measured on the merged branch or after review fixes. Static Lift success markers now resend positions to support late backend activation; the earlier marker-index-only optimization is not included. Runtime benchmark (isaaclab benchmark runtime), Newton/MJWarp, seed 42, no visualizer. Three fresh processes per revision, run in alternating order. Figures are the mean step time across processes; every workload's per-process ranges are disjoint.

Workload	`develop`	This PR	Change
`Isaac-Ant`, 4096 envs	3.369 ms	2.629 ms	−22.0% (+28.1% throughput)
`Isaac-Open-Drawer-Franka-Direct`, 4096 envs	11.106 ms	10.100 ms	−9.1% (+10.0%)
`Isaac-Reorient-Cube-Allegro-Direct`, 4096 envs	27.114 ms	25.221 ms	−7.0% (+7.5%)
`Isaac-Velocity-Rough-AnymalD`, 4096 envs	23.078 ms	22.573 ms	−2.2% (+2.2%)
`Isaac-Lift-KukaAllegro-Camera`, 256 envs, `presets=duo_camera,rgb128`, Newton renderer	12.396 ms	12.202 ms	−1.6% (+1.6%)

Per-process step times (develop → PR):

Ant: 3.367, 3.359, 3.382 → 2.639, 2.638, 2.611 ms.
Drawer: 11.113, 11.102, 11.104 → 10.099, 10.100, 10.100 ms.
Reorient: 27.136, 27.136, 27.070 → 25.217, 25.250, 25.197 ms.
AnymalD: 23.081, 23.050, 23.103 → 22.551, 22.568, 22.599 ms.
Camera: 12.389, 12.374, 12.424 → 12.373, 12.119, 12.113 ms.

Host synchronizations per env.step(), attributed to call sites with PyTorch sync-debug mode over 170 steps:

Task	`develop`	This PR	Remaining
`Isaac-Velocity-Rough-AnymalD`	6.75	3.00	step reset, command resample, and interval-event `nonzero`
`Isaac-Lift-KukaAllegro-Camera`	2.11	2.04	step reset and command resample `nonzero`; curriculum difficulty `.item()` (0.04 per step)

PyTorch labels sync-debug mode a prototype that does not detect every synchronization, and it does not see synchronizations issued by Warp, so these counts are lower bounds.

Validation

The shared sampling refactor passed 257 tests across the existing math, root-state event, and Lift configuration suites. The existing event test now checks exact seeded samples, key order, missing axes, fresh draws, range edits, and warm-cache CUDA synchronization. Eight standalone equivalence checks also passed across CPU/CUDA, float32/float64, and flat/nodal batch shapes. No new tests or test files were added.
Before the shared-sampler refactor, 205 focused tests passed, with each file run in its own process: curriculum, observation, reward, and termination managers; heading commands and mutable event ranges; DiffIK; wrench composition; ray-caster and sensor-base behavior; marker visualization; Lift and reorientation helpers; and the existing joint-limit tests on both PhysX and OvPhysX, on CPU and CUDA. Kit tests used the isaacsim extra, and OvPhysX used ovphysx with separate CPU/GPU processes.
Regression checks fail with the old curriculum accessor, stale drift, and static marker behavior restored. Mutations also turn the retained tests red for custom-noise aliasing, eager debug-visualization refreshes, partial reward reset corruption, and missing reorientation NaN statistics.
Applied the test-audit skill: seven standalone tests were consolidated into existing reward, wrench, reorientation, marker, ray-caster, sensor-update, and Lift tests. Synchronization tests run on CUDA instead of repeating the same bookkeeping on CPU. No new test functions or files were added in the review-fix commit.
Retained unique heading, curriculum snapshot, mutable-range, and ANYmal symmetry coverage: there is no equivalent existing owner test. The symmetry test checks the new permutation tables against the previous independent index-assignment formulation.
uv run isaaclab -f passed, including changelog checks against the actual PR base, and git diff --check passed.
The full test suite and training smoke runs have not been repeated on this revision; benchmark figures above describe the earlier measured revision.

Release backport

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

Checklist

I have read and understood the contribution guidelines.
I have run the pre-commit checks with uv run --frozen isaaclab -f.
I have made corresponding changes to the documentation.
I have added tests that prove the behavior and checked the regressions against the unfixed code.
I have added a changelog fragment for every touched source package.
My name already exists in CONTRIBUTORS.md.

- Log command, termination, curriculum and success-rate episode metrics as device scalars instead of host floats. - Select heading-controlled envs in UniformVelocityCommand with torch.where instead of nonzero(). - Cache event range tensors by value, and make reset_root_state_uniform honor the ranges passed at call time. - Replace the DiffIK zero-quaternion guard with a sync-free torch.where. - Batch RewardManager per-term updates and resets over a term matrix. - Skip defensive copies around built-in noise and delay buffers. - Shift small CircularBuffer frames in two kernels regardless of length. - Check event modes via dict membership in the env step.

- Skip ray-caster drift sampling on reset when drift ranges are zero and upload ray-cast drift ranges only when they change. - Stop eager per-step refreshes for debug-visualized sensors; debug callbacks now refresh outdated buffers lazily before drawing. - Return early from VisualizationMarkers.visualize when no backend is active. - Drop per-frame channel-compacting copies in the Isaac RTX renderer by indexing padded annotator tiles with their source channel count. - Remove the Newton Warp renderer read_output copy loop (outputs alias the camera buffers) and an unreachable PhysX ray-caster guard.

…path

…t-path

Store direct-task joint, body, and tendon ids as device tensors, precompute static limits and default poses, use the actuator target-command API, log success metrics as 0-d tensors, compute Lift contact terms from one stacked norm, cache the ANYmal symmetry permutations and constant tensors, and send only marker indices for static success markers.

…-path

StafaH · 2026-09-28T05:43:01Z

run-ci

…perf/camera-perf-07b-hot-path # Conflicts: # source/isaaclab_tasks/test/core/test_lift_env_cfg.py

StafaH · 2026-09-28T05:47:49Z

run-ci

greptile-apps · 2026-09-28T05:48:01Z

Confidence Score: 4/5

[Medium risk] Performance optimization removing host synchronizations from hot paths.

The PR should not merge until resetting a ray caster after drift is disabled clears its previously sampled offsets.

Findings

Disabled drift remains active ▶
Late markers lack positions ▶

Summary

The PR removes device readbacks and repeated work from camera-adjacent simulation hot paths, batches reward accounting, and updates sensor, backend, and task behavior.

Fixes elapsed-time scaling for body acceleration and makes root-state reset ranges respond to call-time changes.
Adds regression and equivalence tests for the affected managers, sensors, renderers, and tasks.

_{Reviews (1) · Last reviewed commit: "Merge branch 'perf/hot-path-7b-tasks' in..."}

greptile-apps · 2026-09-28T05:48:04Z

+        if not (sample_drift or sample_ray_cast_drift):
+            return


Disabled drift remains active

If a ray caster’s drift range changes from nonzero to zero, reset() leaves the previously sampled drift in its buffer. Later ray casts still use that offset, so disabling drift does not stop it from perturbing sensor measurements. Clear the selected drift buffer when its range becomes zero.

Knowledge Base Used: Simulation, rendering, and sensors

Fixed in ec65628: when a drift range becomes zero, reset() now zeroes the corresponding buffer instead of skipping it (the skip only applies while nothing has ever been sampled).

greptile-apps · 2026-09-28T05:48:04Z

+        if self._static_success_vis_pos_w is not None:
+            # positions and environment ids were set at construction; resending them copies to the host
+            self.success_visualizer.visualize(marker_indices=success.int())
+            return


Late markers lack positions

If the lift command is created before a marker backend is active, its one-time position update is discarded. When a backend becomes active later, these calls send only marker_indices, so the success markers never receive their per-environment positions. Resending the positions on first backend activation would preserve the markers in that case.

Knowledge Base Used: Simulation, rendering, and sensors

Fixed in ec65628: the static-marker shortcut is removed, so the success markers always receive their positions and environment ids on every visualize call and a late-activating backend gets them too.

isaaclab-review-bot

Isaac Lab Review Bot

The hot-path optimizations are broadly coherent, but CurriculumManager.get_active_iterable_terms() now returns cloned tensors while its public type contract still promises floats. Align this accessor’s implementation, annotation, documentation, and consumers before merge.

Design and architecture: The device-resident logging and caching strategy is consistent with the performance goal. The remaining concern is that a reset-path synchronization optimization was also applied to a public inspection accessor without defining the resulting tensor-valued contract.
API: CurriculumManager.get_active_iterable_terms() is declared as returning Sequence[tuple[str, Sequence[float]]], but changed tensor states from Python floats to 0-d tensors. Either retain .item() in this accessor or explicitly update its annotation, documentation, changelog guidance, and affected consumers.
Implementation: At curriculum_manager.py:165 and the equivalent scalar-state path, detach().clone() deterministically preserves tensor values rather than producing the floats promised by the signature. The implementation and public contract therefore need to be reconciled.

Minor fixes needed. Posted 1 actionable finding inline.

Automated review; human maintainers own approval decisions.

AntoineRichard · 2026-09-30T07:21:10Z

We're allocating new tensors here, we could avoid that if we pre-allocated

AntoineRichard · 2026-09-30T07:26:49Z

-    """Per-environment contact force magnitude [N] of a single-body, single-filter contact sensor."""
-    force = sensor.data.normal_force_matrix_w.torch.view(num_envs, 3)
-    return torch.linalg.norm(force, dim=-1)
+def _contact_force_mags(env: ManagerBasedRLEnv, sensor_names: Sequence[str]) -> torch.Tensor:


_contact_force_mag made more sense?

StafaH · 2026-09-30T16:49:22Z

run-ci

… uv docs CLI (#8118, #8105, #8179) (#8210) ## Summary Backports three `develop` PRs to `release/3.0.0`, in merge order: 1. #8118 — towncrier changelog tooling 2. #8105 — camera hot-path cleanup [7B/7] 3. #8179 — uv CLI for Windows docs builds The automatic backport failed for #8118 and #8105. For #8179, this supersedes the bot's draft: - #8209 — bot draft backport of #8179 Once merged, the nightly compiles `release/3.0.0` again: it skips release branches without `tools/changelog/towncrier.toml`. # Description Each PR is cherry-picked with `-x`; every code file matches its source patch. Release-specific changes: 1. #8118 (`ee99db6a820`) - Each `CHANGELOG.rst` conflict resolved as the marker line above the release entries. - `uv.lock` unchanged: release's local package pins already match their `pyproject.toml`. - The bot failed here because `uv.lock` pushed its conflict context past the 1.5M-character inference limit. 2. #8105 (`cd917691992`) - Drops the `.github/workflows/build.yaml` hunk: it edits the `performance-smoke` job, which `release/3.0.0` does not have. The bot's validation rejected this omission. - Splits its 10 old-format fragments into the new format with `cli.py`'s `split_legacy`, keeping all 34 bullets; #8118's check rejects the old format. 3. #8179 (`c1a8bfa50b5`) - Deletes `docs/make.bat`; release held an older copy. - Splits its fragment the same way. - Release's `docs/_redirect/index.html` still hardcodes `./main/` (#7767 is not on this branch), so `isaaclab --docs_multi` writes that redirect unchanged, as release's `make multi-docs` does. ## Type of change - Release backport, carrying #8105's breaking change (`isaaclab` major bump) - New feature, bug fix and documentation update from the source PRs ## Validation 1. `cli.py check release/3.0.0` passes. 2. `tools/changelog` tests: 18 passed. 3. `pre-commit` over the range passes. 4. `cli.py compile` on a scratch copy of this branch: - 9 packages bumped, including `isaaclab` 25.1.0 → 26.0.0 and `isaaclab_tasks_experimental` 0.0.1 → 1.0.0 - every fragment consumed, only its outputs staged - `uv.lock` changed in version lines only 5. Runtime coverage for #8105 and #8179 is left to CI. ## Release backport This PR targets the release branch; no additional backport is needed. ## Checklist - [x] I have read and understood the [contribution guidelines](https://isaac-sim.github.io/IsaacLab/main/source/refs/contributing.html) - [x] I have run the [`pre-commit` checks](https://pre-commit.com/) on the changed files - [x] I have made corresponding changes to the documentation (carried from the source PRs) - [ ] My changes generate no new warnings - [x] I have validated the changed behavior with focused coverage (source PRs' tests carried unchanged) - [x] I have added a changelog fragment under `source/<pkg>/changelog.d/` for every touched package (carried from the source PRs; #8105 and #8179 converted to the new format) - [x] I have added my name to the `CONTRIBUTORS.md` or my name already exists there --------- Co-authored-by: Mustafa H <34825877+StafaH@users.noreply.github.com>

## Description The ANYmal symmetry augmentation used by the `rsl_rl_with_symmetry` agents hard-coded joint indices for the PhysX joint order. The velocity tasks default to Newton, where ANYmal's native order is depth-first, so every mirrored joint observation and action came from the wrong joint. This PR resolves the left-right and front-back permutations and signs from the robot's `joint_names`, using the ANYmal `<L|R><F|H>_<HAA|HFE|KFE>` naming, and feeds them into the cached permutation tables from #8105 (now keyed by joint order). The sign conventions are unchanged: HAA is negated for left-right, and HFE/KFE for front-back. With the PhysX order the tables equal the previous ones. Resolving by name matches the ordering contract, where public order is backend-native unless `joint_ordering` is set, and works for any `joint_ordering`; ANYmal-B/C use the same joint names. Pinning `joint_ordering="physx"` in the task would add a per-step reorder on Newton and change the observation and action layout of existing Newton checkpoints. Fixes #8204 Verification on `Isaac-Velocity-Flat-AnymalD` and `Isaac-Velocity-Rough-AnymalD`, on both PhysX and Newton, at the current head: - Wrong mirrored joints go from 12/12 to 0/12 on Newton, and stay at 0/12 on PhysX. - The physical mirror check (base-frame feet of the mirrored configuration against the mirrored feet) goes from 0.17-0.42 m to below 1e-5 m on Newton. - `test_anymal_symmetry.py` keeps the PhysX reference check from #8105 and adds one check that the Newton order is mirrored like the PhysX order with the joint columns reordered; it fails on develop and passes with this change. - The per-call path is unchanged: one cached gather and multiply per transform. ## Type of change - Bug fix (non-breaking change which fixes an issue) ## Release backport - [x]  Backport this pull request to the active release branch after it merges into `develop` ## Checklist - [x] I have read and understood the [contribution guidelines](https://isaac-sim.github.io/IsaacLab/main/source/refs/contributing.html) - [x] I have run the [`pre-commit` checks](https://pre-commit.com/) with `uv run isaaclab --format` - [ ] I have made corresponding changes to the documentation - [x] My changes generate no new warnings - [x] I have validated the changed behavior with focused coverage and applied the [test-audit criteria](https://github.com/isaac-sim/IsaacLab/blob/develop/skills/developer/test-audit/SKILL.md) to any test changes - [x] I have added a changelog fragment under `source/<pkg>/changelog.d/` for every touched package (do **not** edit `CHANGELOG.rst` or bump package versions — CI handles that) - [x] I have added my name to the `CONTRIBUTORS.md` or my name already exists there --------- Co-authored-by: Kelly Guo <kellyg@nvidia.com>

## Description The ANYmal symmetry augmentation used by the `rsl_rl_with_symmetry` agents hard-coded joint indices for the PhysX joint order. The velocity tasks default to Newton, where ANYmal's native order is depth-first, so every mirrored joint observation and action came from the wrong joint. This PR resolves the left-right and front-back permutations and signs from the robot's `joint_names`, using the ANYmal `<L|R><F|H>_<HAA|HFE|KFE>` naming, and feeds them into the cached permutation tables from #8105 (now keyed by joint order). The sign conventions are unchanged: HAA is negated for left-right, and HFE/KFE for front-back. With the PhysX order the tables equal the previous ones. Resolving by name matches the ordering contract, where public order is backend-native unless `joint_ordering` is set, and works for any `joint_ordering`; ANYmal-B/C use the same joint names. Pinning `joint_ordering="physx"` in the task would add a per-step reorder on Newton and change the observation and action layout of existing Newton checkpoints. Fixes #8204 Verification on `Isaac-Velocity-Flat-AnymalD` and `Isaac-Velocity-Rough-AnymalD`, on both PhysX and Newton, at the current head: - Wrong mirrored joints go from 12/12 to 0/12 on Newton, and stay at 0/12 on PhysX. - The physical mirror check (base-frame feet of the mirrored configuration against the mirrored feet) goes from 0.17-0.42 m to below 1e-5 m on Newton. - `test_anymal_symmetry.py` keeps the PhysX reference check from #8105 and adds one check that the Newton order is mirrored like the PhysX order with the joint columns reordered; it fails on develop and passes with this change. - The per-call path is unchanged: one cached gather and multiply per transform. ## Type of change - Bug fix (non-breaking change which fixes an issue) ## Release backport - [x]  Backport this pull request to the active release branch after it merges into `develop` ## Checklist - [x] I have read and understood the [contribution guidelines](https://isaac-sim.github.io/IsaacLab/main/source/refs/contributing.html) - [x] I have run the [`pre-commit` checks](https://pre-commit.com/) with `uv run isaaclab --format` - [ ] I have made corresponding changes to the documentation - [x] My changes generate no new warnings - [x] I have validated the changed behavior with focused coverage and applied the [test-audit criteria](https://github.com/isaac-sim/IsaacLab/blob/develop/skills/developer/test-audit/SKILL.md) to any test changes - [x] I have added a changelog fragment under `source/<pkg>/changelog.d/` for every touched package (do **not** edit `CHANGELOG.rst` or bump package versions — CI handles that) - [x] I have added my name to the `CONTRIBUTORS.md` or my name already exists there --------- Co-authored-by: Kelly Guo <kellyg@nvidia.com> (cherry picked from commit abe7db4)

StafaH added 12 commits September 27, 2026 22:03

Skip wrench composer resets while inactive

2300758

Remove redundant solver reset from Newton physics step

f0fb3fb

Use elapsed time for Newton body COM acceleration

bc225bf

Read back clamped joint defaults only when logged

5bd37c1

Merge branch 'perf/hot-path-7b-newton' into perf/camera-perf-07b-hot-…

f6b9771

…path

Merge branch 'perf/hot-path-7b-managers' into perf/camera-perf-07b-ho…

5ccec99

…t-path

Merge branch 'perf/hot-path-7b-sensors' into perf/camera-perf-07b-hot…

2792d5f

…-path

Use elapsed time for OVPhysX rigid-object COM acceleration

0c5612a

Merge branch 'perf/hot-path-7b-tasks' into perf/camera-perf-07b-hot-path

8eccb93

StafaH requested a review from a team September 28, 2026 05:42

StafaH requested review from AntoineRichard, ClemensSchwarke, fatimaanes, huidongc, hujc7, kellyguo11, marcodiiga, ooctipus, pbarejko and rilei-nvidia as code owners September 28, 2026 05:42

github-actions Bot added the isaac-lab Related to Isaac Lab team label Sep 28, 2026

isaaclab-bot Bot added ci:run-docker Trigger the on-demand Docker and GPU CI workflow and removed ci:run-docker Trigger the on-demand Docker and GPU CI workflow labels Sep 28, 2026

Merge branch 'develop' of https://github.com/isaac-sim/IsaacLab into …

7b5b894

…perf/camera-perf-07b-hot-path # Conflicts: # source/isaaclab_tasks/test/core/test_lift_env_cfg.py

greptile-apps Bot reviewed Sep 28, 2026

View reviewed changes

Trim redundant hot-path regression tests

282a75b

isaaclab-review-bot Bot reviewed Sep 28, 2026

View reviewed changes

Comment thread source/isaaclab/isaaclab/managers/curriculum_manager.py Outdated

AntoineRichard approved these changes Sep 28, 2026

View reviewed changes

StafaH and others added 8 commits September 29, 2026 09:22

Address hot-path review feedback and consolidate tests

ec65628

Centralize sampling from named event ranges

a71c9af

Require noise callbacks to preserve their inputs

e8a0e12

Keep noise configuration docs concise

cd13a32

Trim redundant host-sync comments

e81da34

Merge branch 'develop' into perf/camera-perf-07b-hot-path

e300905

Remove sync-focused curriculum and velocity command tests

ea03b65

Trim sync, RNG, and private-state test assertions

dc82bb9

AntoineRichard approved these changes Sep 30, 2026

View reviewed changes

Restore contact force magnitude helper name

1ca362a

isaaclab-bot Bot added ci:run-docker Trigger the on-demand Docker and GPU CI workflow and removed ci:run-docker Trigger the on-demand Docker and GPU CI workflow labels Sep 30, 2026

Fix hot-path PR test failures and benchmark cache permissions

e9d4425

StafaH requested a review from myurasov-nv as a code owner September 30, 2026 17:22

github-actions Bot added the infrastructure label Sep 30, 2026

StafaH added the ci:run-docker Trigger the on-demand Docker and GPU CI workflow label Sep 30, 2026

ooctipus approved these changes Sep 30, 2026

View reviewed changes

ooctipus merged commit cd91769 into isaac-sim:develop Sep 30, 2026
54 of 55 checks passed

ooctipus deleted the perf/camera-perf-07b-hot-path branch September 30, 2026 18:46

NeoZng mentioned this pull request Sep 30, 2026

Fix ANYmal symmetry augmentation for non-PhysX joint orders #8205

Merged

7 of 8 tasks

hujc7 mentioned this pull request Sep 30, 2026

[Backport release/3.0.0] Towncrier changelog, camera hot-path 7B, and uv docs CLI (#8118, #8105, #8179) #8210

Merged

6 of 7 tasks

ooctipus mentioned this pull request Sep 30, 2026

[Core] Resolve manager term defaults and validate randomization distributions #8127

Merged

8 tasks done

Conversation

StafaH commented Sep 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Changes

Host synchronizations removed

Launch overhead

Overengineered code removed

Bug fixes

Investigated and not included

Type of change

Performance

Validation

Release backport

Checklist

Uh oh!

StafaH commented Sep 28, 2026

Uh oh!

StafaH commented Sep 28, 2026

Uh oh!

greptile-apps Bot commented Sep 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Confidence Score: 4/5

Findings

Summary

Uh oh!

greptile-apps Bot Sep 28, 2026

Choose a reason for hiding this comment

Uh oh!

StafaH Sep 30, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps Bot Sep 28, 2026

Choose a reason for hiding this comment

Uh oh!

StafaH Sep 30, 2026

Choose a reason for hiding this comment

Uh oh!

isaaclab-review-bot Bot left a comment

Choose a reason for hiding this comment

Isaac Lab Review Bot

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

AntoineRichard Sep 30, 2026

Choose a reason for hiding this comment

Uh oh!

AntoineRichard Sep 30, 2026

Choose a reason for hiding this comment

Uh oh!

StafaH commented Sep 30, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

StafaH commented Sep 28, 2026 •

edited

Loading

greptile-apps Bot commented Sep 28, 2026 •

edited

Loading