Skip to content

cudanav_gpu_closed_loop_release_2026-07-29.md

Latest commit

 

History

History
56 lines (43 loc) · 1.93 KB

cudanav_gpu_closed_loop_release_2026-07-29.md

File metadata and controls

56 lines (43 loc) · 1.93 KB
Raw

CudaNav native all-GPU 10-minute closed-loop release

Date: 2026-07-29

Source commit: b2c490799ad098c5cc0f26127c18c8d8d186ccf2

Profile: release

Result: PASS

CudaNav native all-GPU closed-loop release

The deterministic S-course plant generates LiDAR from ground truth, but the controller only receives the GPU KISS-ICP estimate. Each CUDA MPPI command is applied to the plant before the next scan. GPU voxel mapping and GPU ESDF build the controller costmap in the same process.

Result

  • Goal reached: true
  • Traversals: 30/30
  • Simulated duration: 1059.4 s
  • Final ground-truth goal distance: 0.296 m
  • Collision count: 0
  • Ground-truth distance: 352.748 m
  • Command-effect distance: 352.748 m
  • KISS-ICP ATE RMSE: 0.012 m
  • KISS-ICP final drift: 0.003%
  • Minimum ICP inliers: 213
  • Final observed voxels: 53012
  • Peak occupied 2D cells: 760
  • MPPI solve p95: 0.455 ms
  • Full frame p95: 5.237 ms
  • Command deadline miss rate: 0.000%
  • All-colliding evaluations: 0
  • Minimum nonzero valid-rollout ratio: 0.195

Visual evidence

The 180-frame GIF retains the start and end of every traversal and is bound to the 10,594-row trajectory by SHA-256. Its machine-readable sidecar records the source evidence, trajectory, renderer, dimensions, frame count, and GIF hash: cudanav_gpu_closed_loop_release.json.

python scripts/render_cudanav_gpu_closed_loop.py \
  --evidence docs/results/cudanav_gpu_closed_loop_release_2026-07-29.json \
  --trajectory build/cudanav_gpu_closed_loop_release_b2c4907/trajectory.csv \
  --output gif/cudanav_gpu_closed_loop_release.gif \
  --manifest gif/cudanav_gpu_closed_loop_release.json \
  --check-only

Scope

Native deterministic S-course simulation. CUDA MPPI commands are applied to the plant and affect later LiDAR scans. This is not a ROS 2 runtime result and does not use recorded real-world data.