Batch, steps, dt
`batch_size`, `time_steps`, `iteration_count`, `model_dt`, and `temperature` must be positive.
End-to-End CudaNav Development Smoke
Run the complete command-driven loop from the current source checkout.
docker build --pull --no-cache -f docker/Dockerfile -t cudarobotics .
docker run --rm --gpus all -v "$PWD/out:/out" cudarobotics cudanav
# ROS 2 Jazzy source workspace
ros2 launch cuda_nav_bringup cudanav_closed_loop.launch.pyThe image connects GPU KISS-ICP, rolling voxel mapping, typed ESDF, the Nav2 costmap layer, CUDA MPPI, and the simulator. The command exits non-zero unless `/out/cudanav_closed_loop.json` passes its short smoke gate. Release evidence still uses the separate 10-minute retained harness.
Build and Smoke Test
cd ros2_ws
colcon build --packages-select cuda_mppi_controller \
--cmake-args -DCMAKE_BUILD_TYPE=Release
source install/setup.bash
ros2 run cuda_mppi_controller plugin_load_test
ros2 run cuda_mppi_controller parameter_validation_test
ros2 run cuda_mppi_controller mppi_gpu_standalone 2048
ros2 run cuda_mppi_controller mppi_gpu_standalone 2048 esdf
Controller Server Config
Point `controller_server` at the plugin and set the MPPI horizon, sample count, model, and limits.
controller_server:
ros__parameters:
controller_plugins: ["FollowPath"]
FollowPath:
plugin: "cuda_mppi_controller::CudaMppiController"
batch_size: 8192
time_steps: 56
model_dt: 0.05
motion_model: "DiffDrive"
path_angle_weight: 0.25
curvature_speed_weight: 0.0
curvature_speed_min: 0.18
distance_field_weight: 0.0
distance_field_cutoff: 0.75Full example: cuda_mppi_params.example.yaml.
Validated Parameters
Invalid configuration is rejected during configure and live parameter updates before the optimizer is rebuilt.
Velocity and model limits
Control bounds must be finite and internally consistent for the selected motion model.
Weights and lookahead
Weights, distance-field cutoff, lookahead distances, transform tolerance, and retreat scale reject invalid values.
Motion Models
| Model | Controls | Notes |
|---|---|---|
| DiffDrive | `vx`, `wz` | Default model for differential-drive bases. |
| Ackermann | `vx`, curvature-limited `wz` | Uses `min_turning_r` to limit angular rate by forward speed. |
| Omni | `vx`, `vy`, `wz` | Adds lateral velocity sampling and limits. |
Curvature Speed Critic
Set `curvature_speed_weight` above zero to slow rollout speed near sharp path bends.
The critic estimates local path curvature around the follow point and penalizes forward speed above a curvature-limited target with `curvature_speed_min` as the floor. It stays disabled by default because it is a bend-entry smoothness knob rather than a universal time-to-goal improvement.
Distance Field Critic
Enable `distance_field_weight` to add a smooth clearance cost near obstacles without changing the default costmap-only behavior.
The controller builds a truncated GPU distance-to-obstacle field from the local costmap each cycle. Cells beyond `distance_field_cutoff` receive no extra penalty, while rollouts near lethal or inscribed cells receive a quadratic clearance cost.
Diagnostics and Bag Evaluation
Capture per-cycle solve, validity, cost, retreat, and command signals when moving beyond synthetic benchmarks.
Diagnostics CSV
Set `diagnostics_csv_path`, then render the result with `scripts/render_cuda_mppi_diagnostics.py`.
python3 scripts/render_cuda_mppi_diagnostics.py \
/tmp/cuda_mppi_diagnostics.csv \
--output-stem docs/results/cuda_mppi_diagnostics_runBag / real-data harness
Use `scripts/run_cuda_mppi_bag_eval.py` to coordinate Nav2 startup, rosbag playback, topic recording, and diagnostics rendering.