Model Predictive Path Integral Controller

CUDA-accelerated MPPI controller based on nav2_mppi_controller built on the MPPI-Generic CUDA framework.

For ROS1, see the noetic branch.

For ROS2-Jazzy, see the jazzy branch.

Installation

Requires an NVIDIA GPU + CUDA toolkit. Tested with ROS2 Humble, Ubuntu 22.04, CUDA 12.6, CPU AMD Ryzen 5 5500H (8 threads) + RTX 2050 (compute capability 8.6); Jetson AGX Orin Developer Kit (JetPack 6.2.3, Ubuntu 22.04.5, CUDA 12.6, compute capability 8.7).

• Install the CUDA toolkit (see NVIDIA's install guide).
• Clone and build:

  cd your_ws/src
  git clone -b humble https://github.com/datledoan/mppi_controller_cuda.git
  cd your_ws
  colcon build --packages-select mppi_controller_cuda --parallel-workers 2

  CMAKE_CUDA_ARCHITECTURES defaults to 86 (desktop Ampere, e.g. RTX 2050). Override for other GPUs, e.g. 87 for Jetson Orin (AGX Orin/Orin NX/Orin Nano):

  colcon build --packages-select mppi_controller_cuda --parallel-workers 2 --cmake-args -DCMAKE_CUDA_ARCHITECTURES=<your-arch>

Usage

Set the plugin type under your controller_server's FollowPath in your Nav2 params yaml:

controller_server:
  ros__parameters:
    controller_plugins: ["FollowPath"]
    FollowPath:
      plugin: "mppi_controller_cuda::MPPIControllerCUDA"
      # ... see params/mppi_controller_params.yaml for every parameter + its default

See params/mppi_controller_params.yaml for the full list with defaults. Example with turtlebot3 burger: turtlebot3

Result

Demo

Benchmark

Real-run controller_server CPU usage while navigating a goal in Gazebo (turtlebot3 burger, see turtlebot3), sampled with benchmark/run_live_cpu_test.sh (10s window, 0.2s interval, 5 fresh-relaunch runs x 50 samples each):

Controller	CPU usage (of 1 core, mean ± std across 5 runs)
nav2_mppi_controller (CPU)	30.9% ± 0.9
mppi_controller_cuda (GPU)	9.4% ± 0.5

~69% less controller_server CPU load with the GPU plugin (RTX 2050).

Same setup on a Jetson AGX Orin Developer Kit (JetPack 6.2.3, Ubuntu 22.04.5, CUDA 12.6, built with -DCMAKE_CUDA_ARCHITECTURES=87 for Orin's compute capability 8.7) -- Gazebo running on a separate host PC, controller_server on the Jetson over the network:

Controller	CPU usage (of 1 core, mean ± std across 5 runs)
nav2_mppi_controller (CPU)	65.4% ± 0.8
mppi_controller_cuda (GPU)	18.6% ± 0.1

~72% less controller_server CPU load with the GPU plugin (Jetson AGX Orin).

Per-call compute time (Optimizer::evalControl()/Controller::computeControl() -- see benchmark/), swept across batch_size/time_steps:

Reference

• nav2_mppi_controller
• MPPI-Generic

@misc{vlahov2024mppi,
      title={MPPI-Generic: A CUDA Library for Stochastic Trajectory Optimization},
      author={Bogdan Vlahov and Jason Gibson and Manan Gandhi and Evangelos A. Theodorou},
      year={2024},
      eprint={2409.07563},
      archivePrefix={arXiv},
      primaryClass={cs.MS},
      url={https://arxiv.org/abs/2409.07563},
}