LLA-MPC on Embedded Hardware: Rapid Adaptive Control with Thousands of Parallel Models

Can a learning-free, parallel model-bank MPC identify rapidly changing tire parameters online and control a small-scale autonomous car on embedded hardware under noisy state estimation?

Model-based controllers fail when vehicle dynamics change rapidly—worn tires, surface changes, shifting payloads—because their fixed model becomes inaccurate. Existing adaptation often needs training data, warm-up laps, model derivatives, or continuous parameter updates that cannot jump quickly to a true model after an abrupt change.

Prior methods include classical multiple-model adaptive control switching among controllers, multiple-model predictive control blending linear models by residuals, adaptive MPC estimating parameters continuously, and online learning approaches such as Gaussian-process residuals or meta-learned models pretrained offline. These either require pre-designed controllers, offline training, warm-up data, derivatives, or blending that cannot switch to a distant model quickly.

The authors generalize LLA-MPC with modular dynamics, integrators, and parallel backends. The controller holds a bank of thousands of nonlinear single-track bicycle models with Fiala brush tire forces, varied over cornering stiffness and friction parameters. At each step it rolls all models forward in parallel with JAX, computes one-step residuals over a sliding window of 60 transitions, selects the best model by quadratic error, and passes it to an MPC solver with a 30-step horizon. The code runs on both an Intel laptop and a Jetson Orin Nano; the Orin GPU evaluates at least 15,000 model rollouts at 40 Hz.

Two main configurations were tested: offboard Intel 258V with motion-capture pose at 180 Hz and up to 40 Hz control, and onboard Jetson Orin Nano with particle-filter pose at 40 Hz. In low-friction tracking, LLA-MPC tracked for over 8 laps while the nominal controller crashed on the third lap. Across changing surfaces with an abrupt height change, LLA-MPC completed over 6 laps; nominal completed only one lap. In an onboard fast-turn task, LLA-MPC completed a high-speed drifting maneuver where nominal crashed after a wide arc. LLA-MPC estimated lower friction on low-friction ground and higher friction on higher-grip foam, and had lower H-step prediction error than nominal in successful runs, though quantitative error values are not reported. The model bank contained 5000 models, with window size W=60 and horizon H=30.

The paper is not peer reviewed. Evaluation is on a single F1TENTH platform with plastic low-friction tires and a small number of configurations; no statistically repeated trials or explicit success-rate percentages are reported. Quantitative H-step prediction error values are not provided. Authors note future work: weighting look-back timesteps by uncertainty and excitation, systematic hyperparameter selection, and integration with planning, safety, and learning. Validation is small-scale and does not yet cover full-size vehicles or safety-critical certification; transfer beyond this car and lab surfaces is unproven.