mppi

Standalone MPPI control library extracted from BeamNGRL.control.UW_mppi.

Contents

• mppi.MPPI — Williams-style MPPI shell
• mppi.Sampling.Delta_Sampling — delta-control sampling (trailing control_dim state channels)
• mppi.Dynamics.SimpleCarDynamicsTCUDA — CUDA bicycle + BEV (analytical_bicycle.{cu,cpp} next to the module)
• mppi.Costs — example pure-torch costs (not used by HOUND deployment)
  • SimpleCarCost — goal-seeking (set_goal / set_BEV)
  • TrackingCost — path tracking (set_path / set_BEV)
• mppi/Configs/*.yaml — sample Dynamics / Sampling / Cost / Map / MPPI configs

HOUND deployment tracking cost is hound_nav.trackingCostCUDA.SimpleCarCost (CUDA). Swap that in the same MPPI(Dynamics, Costs, Sampling, …) constructor.

Cost API (for custom costs)

forward(states, controls) -> [K]     # M×K×T×NX states, M×K×T×nu controls
set_BEV(height, normal, cost)        # optional
set_goal(xy) / set_path(T×4)         # optional

Example

# after colcon build + source install/setup.bash
python3 /path/to/mppi/examples/wire_mppi_cost.py --cost tracking
python3 /path/to/mppi/examples/wire_mppi_cost.py --cost goal

NX / plant size

Pass Dynamics_config["NX"] (or state_dims / control_state_dims). Default and minimum is 17 (indices 0..14 used by the bicycle kernel; last 2 = actuators).

Build

cd ~/colcon_ws && colcon build --packages-select mppi hound_nav --symlink-install