NonPlanarMPC

Reference controller for Nonplanar Model Predictive Control for Autonomous Vehicles with Recursive Sparse Gaussian Process Dynamics (Amine, Puri, Le, Mangharam — IEEE Intelligent Vehicles Symposium, 2026).

The method is a sampling-based MPPI controller on a single-track dynamic bicycle model, augmented with an online-learned recursive sparse Gaussian-process residual that compensates for nonplanar terrain (slopes, banks, hills).

Important

The controller implementation is being finalized (cleanup) and will be released here shortly. For early access or questions in the meantime, please contact Ahmad Amine (aminea at upenn.edu).

Simulation environment

The custom NVIDIA Isaac Sim environment originally developed for this project is deprecated. The simulation project evolved into the open-source Autoware off-road sim, included here as a submodule under third_party/. It ships the nonplanar tracks used in the paper (L-shaped, kidney, oval) and a RoboRacer vehicle with a ROS 2 interface. All credits for the sim go to the original authors of the Autoware off-road sim. The NonPlanarMPC controller is compatible with the sim and can be run on the tracks with the launch configurations in sim_configs/.

Repository layout

Path	Contents
third_party/autoware_off-road_sim/	Simulation environment - submodule
third_party/f1tenth_planning/	Motion-planning library the controller builds on - submodule
sim_configs/	Off-road sim launch configs - coming soon
tracks/	Per-track raceline + terrain data - coming soon
scripts/	Run / training / extraction entry points - coming soon
docs/	Project website - coming soon

The npmpc controller package will be added at the top level when released.

Submodules

git clone --recurse-submodules https://github.com/mlab-upenn/NonPlanarMPC.git
# or, after a plain clone:
git submodule update --init --recursive

Citation

@misc{amine2026nonplanarmodelpredictivecontrol,
      title={Nonplanar Model Predictive Control for Autonomous Vehicles with Recursive Sparse Gaussian Process Dynamics}, 
      author={Ahmad Amine and Kabir Puri and Viet-Anh Le and Rahul Mangharam},
      year={2026},
      eprint={2602.16206},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2602.16206}, 
}