Skip to content

v2.2.0 - ROS 2 Jazzy GPU correctness and throughput

Latest

Choose a tag to compare

@Idate96 Idate96 released this 20 Jul 11:00
· 3 commits to ros2 since this release

ROS 2 Jazzy GPU correctness and throughput

This release makes the CuPy elevation-mapping path deterministic, corrects geometry and dilation defects, and substantially reduces mapping and GridMap publication latency. It keeps the existing ROS interfaces and does not add NVIDIA Warp as a runtime dependency.

Highlights

  • Correct Euclidean dilation without row wrapping, with a direct small-radius kernel and EDT for large radii.
  • Race-free point fusion and visibility cleanup using immutable snapshots, proposal buffers, and deterministic per-cell finalization.
  • Exact 2D grid DDA visibility traversal, float32 geometry, and early NaN/Inf rejection.
  • Correct organized PointCloud2 parsing with row padding and truncated-buffer validation.
  • Generation-based plugin caching, race-free min filtering, reusable GPU buffers, and CuPy device-pool allocation.
  • Faster GridMap payload encoding and no publication work when a topic has no subscribers.
  • Existing ROS 2 improvements since v2.1.0, including map-center/height-gate fixes, bounded inpainting and despike filters, semantic workflows, and the clear_map service.
  • Reproducible Jazzy CI dependencies by building GridMap 2.2.2 against the pinned container's middleware ABI.

Measured performance

On an NVIDIA RTX 4090, using three independent runs with 30 warmups and 200 measured iterations:

  • Core mapping callback p95 improved by 55.3–64.2% for 10k–100k-point clouds.
  • In-process ROS callback plus six-layer filtered GridMap preparation p95 improved by 26.3–65.8%.
  • 602×602 large-radius dilation improved from 8.704 ms to 0.590 ms p95.
  • 600×600 single-layer GridMap construction improved from 10.80 ms to 0.90 ms p95.

The equivalent finalization algorithm was benchmarked in CuPy and Warp. Warp showed no repeatable p95 benefit, so CuPy RawKernel remains the implementation.

Detailed methodology is in GPU_OPTIMIZATION_WORKLOG.md; machine-readable results are in benchmarks/results/rtx4090_20260720_summary.json.

Validation

  • Exact pinned Jazzy GPU CI container on starship: 10 packages built.
  • Semantic image and pointcloud runtime smokes passed.
  • 77 tests passed with zero errors, failures, or skips.
  • TF/GridMap integration, save/load services, and synthetic demo launch passed.
  • GitHub Sphinx documentation workflow passed.

The repository's remote GPU job was cancelled while queued because no runner is currently registered for this repository. The same pinned image and workflow script were executed locally on the intended RTX 4090 host as the release gate.

Deployment note

A representative raw Ouster bag was not available in the workspace. Before robot deployment, replay a recorded cloud and measure executor, serialization, DDS latency, and publication deadlines on the target system.