Contextual Tuning of Model Predictive Control for Autonomous Racing

TL;DR

The paper addresses the challenge of adapting both a predictive dynamics model and controller parameters in learning-based MPC for autonomous racing under changing conditions. It introduces a residual dynamics model that encodes environmental context and employs contextual Bayesian optimization to transfer knowledge across contexts, enabling data-efficient tuning. The approach is validated on a hardware platform with 1:28 RC cars, showing improved prediction accuracy and faster convergence to context-specific optimal controllers than standard Bayesian optimization. The work demonstrates that context-aware learning and tuning can significantly enhance closed-loop performance in autonomous racing with limited data, and provides an open-source implementation for the community.

Abstract

Learning-based model predictive control has been widely applied in autonomous racing to improve the closed-loop behaviour of vehicles in a data-driven manner. When environmental conditions change, e.g., due to rain, often only the predictive model is adapted, but the controller parameters are kept constant. However, this can lead to suboptimal behaviour. In this paper, we address the problem of data-efficient controller tuning, adapting both the model and objective simultaneously. The key novelty of the proposed approach is that we leverage a learned dynamics model to encode the environmental condition as a so-called context. This insight allows us to employ contextual Bayesian optimization to efficiently transfer knowledge across different environmental conditions. Consequently, we require fewer data to find the optimal controller configuration for each context. The proposed framework is extensively evaluated with more than 3'000 laps driven on an experimental platform with 1:28 scale RC race cars. The results show that our approach successfully optimizes the lap time across different contexts requiring fewer data compared to other approaches based on standard Bayesian optimization.

Figures (7)

• Figure 1: The dynamic bicycle model with corresponding velocities (blue) and forces (red). The inputs are given by the longitudinal drive train command $\tau$ acting on $F_x$ and the steering angle $\delta$. The tire forces $F_{\mathrm{f}/\mathrm{r}}$ are computed with the simplified Pacejka model based on the slip angles $\alpha_{\mathrm{f}/\mathrm{r}}$.
• Figure 2: Schematic overview of the control and learning architecture with their respective communication channels. The subsystems are explained in \ref{['sec:preliminaries', 'sec:learning_architecture']}. The blue arrows denote communication in real-time at 35 Hz for the control loop. The dashed, red arrows denote the communication between the learning components, which occurs once every two laps.
• Figure 3: Error prediction (mean $\pm$ one standard deviation) from the learned residual model with $\bm{\phi}_{\mathrm{Taylor}}$\ref{['eq:pacejka_taylor_approximation']} on training (green) and testing (red) data for lateral velocity (top) and yaw rate (bottom), respectively. The vertical dashed lines indicate the start of a new lap.
• Figure 4: Inferred regression coefficients as a function of the amount of training data (in seconds) for the different contexts (: context 1, : context 2, : context 3, : context 4). The solid lines depict the posterior mean estimate and the shaded area corresponds to mean $\pm$ one standard deviation.
• Figure 5: Three plots on the left-hand side: density of the car's position across 20 laps for three different combinations of controller parameters and dynamic models. Right-hand side: distributions of the corresponding lap times for the three combinations.