CUDA MPPI vs Nav2 CPU MPPI
Closed-loop pluginlib comparison on the same costmap, plan, and limits.
Evidence index
CudaRobotics keeps benchmark reports in the repository so wins, negative controls, and limitations are visible next to the code.
Start here for the published v0.3.0 release evidence.
Closed-loop pluginlib comparison on the same costmap, plan, and limits.
Double-gap and moving-crossing smoke coverage for path-window and dynamic-costmap regressions.
GPU MPPI heading-to-path tangent cost for high-curvature tracking.
Optional path-curvature speed target for smoother bend entry.
GPU-only comparison of default costmap scoring and optional distance-field clearance.
FilterReg comparison against probreg and Open3D CPU baselines.
Five GPU registrars under one schema with explicit quality gates and environment metadata.
Recorded/shadow evidence with command pairing, clearance, and a retained failed quality gate.
Required CPU, GPU, ROS, packaging, and evidence gates before the tag is published.
The fixed-seed suite compares the baseline MPPI planner against a growing set of paper-inspired variants. It is not a universal claim; it is a reproducible local stress suite with both successes and failures checked in.
| Report | Scope |
|---|---|
| 2026-06-10 suite | Five scenarios, ten planners, multiple sample counts. |
| Smoke suite | Small quick-run suite for fast regression checks. |
| Reproduction zoo | Research backlog with implementation notes and limitations. |