Skip to content

Releases: rsasaki0109/Nav2PlannerBattle

v0.3.0

03 Jun 14:45

Choose a tag to compare

Research-driven generative model families + costmap conditioning.

A deep literature/OSS survey found no open-source flow-matching / diffusion / consistency local planner for ROS 2 Nav2 ground robots. This release adds all three behind the TrajectoryModel ONNX seam:

  • Flow matching (single/few-step), diffusion (cosine DDPM + DDIM), consistency (one-step distillation). Each exports a single self-contained ONNX file matching the [1,4] -> [1,K,H,3] backend contract; a trained flow model runs in the real C++ OnnxTrajectoryModel.
  • Costmap+goal conditioning end to end (the surveyed gap): optional egocentric costmap patch through ModelContext → OnnxTrajectoryModel auto-detected costmap input → controller costmap_patch_size (default off, backward compatible). CostmapFlowPlanner / CostmapDiffusionPlanner / CostmapConsistencyPlanner train+export two-input ONNX.
  • README hero GIF from the real pipeline; GPU/headless Gazebo notes.

Verified on ROS 2 Jazzy (build + gtest + pytest + ament lints). See CHANGELOG.md. Not a safety-certified product; see docs/safety.md.

v0.2.0

03 Jun 11:05

Choose a tag to compare

Learned models, end to end.

  • ONNX inference backend (nav2_diffusion_onnx, optional) behind the TrajectoryModel pluginlib seam.
  • Controller model selection at runtime via model_plugin / model_path — no inference lib linked into core/controller.
  • Training pipeline: rosbag / rule-based expert → dataset → PyTorch → ONNX export, matching the backend I/O contract (round trip tested).
  • RViz visualization: candidate markers (best/safe/rejected + rejection text) and SafetyState.
  • Benchmark suite: metrics, safety-first score, leaderboard, aggregation, YAML scenarios, runner.

Verified on ROS 2 Jazzy (colcon build/test, 0 failures). See CHANGELOG.md. Not a safety-certified product; see docs/safety.md.

v0.1.0

03 Jun 07:45

Choose a tag to compare

Nav2-native generative local controller (Controller Plugin / Mode A) with a deterministic safety layer, multimodal candidate generation behind a TrajectoryModel plugin seam, fallback to MPPI/RPP, and a benchmark suite. See CHANGELOG.md.

Verified on ROS 2 Jazzy: colcon build + test of all six packages = 252 tests, 0 failures.

Not a safety-certified product; see docs/safety.md.