Learning-based model predictive control for high-performance miniature racing

With its high demands on driving performance, safety-awareness and computational efficiency, autonomous racing has established itself as a valuable benchmarking application for research on learning-based control algorithms [1]. Thereby, a physics-based model is augmented with a residual model learned from data to enhance prediction accuracy and enable online model adaptation to changing operating conditions, such as temperature, ground surface and tire wear. With respect to the choice of residual model, in particular for online model adaptation, Gaussian process (GP) regression has shown promising results since, by providing a quantification of the modeling uncertainty, it enables to maintain a prescribed level of safety-awareness when incorporating real-world data into the prediction model [2]. Still, especially when considering the learning-based model uncertainty within the control algorithm, the high computational complexity of GP regression constitutes a major bottleneck for real-time applications. In this semester project, you will implement an efficient Gaussian process-based model predictive controller for high-performance autonomous miniature racing.

With its high demands on driving performance, safety-awareness and computational efficiency, autonomous racing has established itself as a valuable benchmarking application for research on learning-based control algorithms [1]. Thereby, a physics-based model is augmented with a residual model learned from data to enhance prediction accuracy and enable online model adaptation to changing operating conditions, such as temperature, ground surface and tire wear. With respect to the choice of residual model, in particular for online model adaptation, Gaussian process (GP) regression has shown promising results since, by providing a quantification of the modeling uncertainty, it enables to maintain a prescribed level of safety-awareness when incorporating real-world data into the prediction model [2]. Still, especially when considering the learning-based model uncertainty within the control algorithm, the high computational complexity of GP regression constitutes a major bottleneck for real-time applications.

In this semester project, you will implement an efficient Gaussian process-based model predictive controller for high-performance autonomous miniature racing.

Your project would include • familiarizing yourself with the existing GP-MPC implementation [3] in the CRS autonomous racing framework, • extending the method by implementing an efficient (approximate) Gaussian process formulation for real-time inference, • performing simulation and real-world experiments to compare your implementation against the existing GP-MPC implementation.

Your project would include

• familiarizing yourself with the existing GP-MPC implementation [3] in the CRS autonomous racing framework,

• extending the method by implementing an efficient (approximate) Gaussian process formulation for real-time inference,

• performing simulation and real-world experiments to compare your implementation against the existing GP-MPC implementation.

Amon Lahr, amlahr@ethz.ch; Dr. Andrea Carron, carrona@ethz.ch

Amon Lahr, amlahr@ethz.ch; Dr. Andrea Carron, carrona@ethz.ch