Comparing Gradient-Based and Sampling-Based Model Predictive Control for Autonomous Racing
(2025)Department of Automatic Control
- Abstract
- This thesis investigates the performance of two model predictive control (MPC) variants for autonomous racing, using a 1:10-scale vehicle as a testbed for reference path-tracking experiments. The platform is equipped with a single inertial navigation system (INS) sensor, providing highly accurate odometry measurements. The controllers under study are the standard gradient-based MPC and a sampling-based approach known as model predictive path integral (MPPI) control. The controllers utilize longitudinally and laterally coupled non-linear kinematic and dynamic bicycle models, and transition between models is achieved through speed-based linear blending of the state update functions. High-fidelity system identification of the vehicle dynamics... (More)
- This thesis investigates the performance of two model predictive control (MPC) variants for autonomous racing, using a 1:10-scale vehicle as a testbed for reference path-tracking experiments. The platform is equipped with a single inertial navigation system (INS) sensor, providing highly accurate odometry measurements. The controllers under study are the standard gradient-based MPC and a sampling-based approach known as model predictive path integral (MPPI) control. The controllers utilize longitudinally and laterally coupled non-linear kinematic and dynamic bicycle models, and transition between models is achieved through speed-based linear blending of the state update functions. High-fidelity system identification of the vehicle dynamics is achieved using a physics-constrained neural network (PCNN), with the learned dynamics employed in simulation for controller tuning. The final evaluation consists of a head-to-head comparison, where the vehicle, driven by each controller in turn, completes a ten-lap time attack on a scaled-down Formula Student Germany 2023 Driverless Cup track. We conclude that MPC outperforms MPPI in terms of lap times and tracking performance, while MPPI exhibits superior consistency in computational time. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9212519
- author
- Meledin, Leonid
- supervisor
- organization
- year
- 2025
- type
- H3 - Professional qualifications (4 Years - )
- subject
- report number
- TFRT-6296
- other publication id
- 0280-5316
- language
- English
- id
- 9212519
- date added to LUP
- 2025-09-18 14:17:07
- date last changed
- 2025-09-18 14:17:07
@misc{9212519,
abstract = {{This thesis investigates the performance of two model predictive control (MPC) variants for autonomous racing, using a 1:10-scale vehicle as a testbed for reference path-tracking experiments. The platform is equipped with a single inertial navigation system (INS) sensor, providing highly accurate odometry measurements. The controllers under study are the standard gradient-based MPC and a sampling-based approach known as model predictive path integral (MPPI) control. The controllers utilize longitudinally and laterally coupled non-linear kinematic and dynamic bicycle models, and transition between models is achieved through speed-based linear blending of the state update functions. High-fidelity system identification of the vehicle dynamics is achieved using a physics-constrained neural network (PCNN), with the learned dynamics employed in simulation for controller tuning. The final evaluation consists of a head-to-head comparison, where the vehicle, driven by each controller in turn, completes a ten-lap time attack on a scaled-down Formula Student Germany 2023 Driverless Cup track. We conclude that MPC outperforms MPPI in terms of lap times and tracking performance, while MPPI exhibits superior consistency in computational time.}},
author = {{Meledin, Leonid}},
language = {{eng}},
note = {{Student Paper}},
title = {{Comparing Gradient-Based and Sampling-Based Model Predictive Control for Autonomous Racing}},
year = {{2025}},
}