Learning How to Autonomously Race a Car: a Predictive Control Approach (1901.08184v4)

Abstract: In this paper we present a Learning Model Predictive Controller (LMPC) for autonomous racing. We model the autonomous racing problem as a minimum time iterative control task, where an iteration corresponds to a lap. In the proposed approach at each lap the race time does not increase compared to the previous lap. The system trajectory and input sequence of each lap are stored and used to systematically update the controller for the next lap. The first contribution of the paper is to propose a LMPC strategy which reduces the computational burden associated with existing LMPC strategies. In particular, we show how to construct a safe set and an approximation to the value function, using a subset of the stored data. The second contribution is to present a system identification strategy for the autonomous racing iterative control task. We use data from previous iterations and the vehicle's kinematics equations to build an affine time-varying prediction model. The effectiveness of the proposed strategy is demonstrated by experimental results on the Berkeley Autonomous Race Car (BARC) platform.

• The paper presents a novel Learning Model Predictive Controller that refines racing performance by iteratively updating safe sets using localized trajectory data.
• The approach employs an Affine Time Varying prediction model and separates planning from tracking to accurately capture vehicle dynamics.
• The study demonstrates computational efficiency with control problems solved in under 10ms, significantly improving over previous implementations.

A Predictive Control Approach to Learning Autonomous Racing

This paper presents a novel Learning Model Predictive Controller (LMPC) specifically designed for autonomous racing applications. The authors, Ugo Rosolia and Francesco Borrelli, propose a control framework wherein a race car learns and optimizes its performance through iterative laps on a predefined track. Broadly, the paper contributes two primary innovations: a localization strategy for model predictive control (MPC) tasks and a tailored method for system identification in racing contexts.

Methodology Overview

The researchers frame the autonomous racing challenge as an iterative control problem, where each iteration represents a full lap completed by the vehicle. Fundamentally, their method centers around the LMPC which consistently improves upon its performance by adjusting both its terminal cost and constraints using a localized approach. This localization drastically reduces computational overhead, compared to full data utilization schemes, by employing subsets of past trajectory data to refine safe sets and approximate value functions.

A significant aspect of the approach is the separation of tasks into higher-level trajectory planning and lower-level trajectory tracking, a theme common in model predictive control frameworks for dynamic applications. The LMPC uses stored data iteratively; it updates the predictive model using both kinematic and dynamic vehicle models to better learn the track and optimize lap times progressively.

Experimental Validation

The practical implications of this research are tested using the Berkeley Autonomous Race Car (BARC) platform. This choice of a physical test-bed is crucial for demonstrating real-world applicability and verifying computational assumptions. The results showed a marked reduction in lap times as the controller iteratively refined its performance. Notably, the vehicle's capability to operate at the edge of its performance envelope—maintaining lateral accelerations close to 1g—was a strong testament to the effectiveness of the proposed approach.

Technical Contributions and Implications

1. Local Safe Sets and Value Function Approximations: By focusing on localized data sets from previous iterations, the paper introduces a method for reducing computational resources while maintaining stringent control performance metrics. The LMPC benefits from a continuous stream of updated localized constraints, allowing it to adapt dynamically to the racetrack conditions.
2. Affine Time Varying (ATV) Prediction Model: This model leverages previous iterations’ data in combination with refined kinematic equations to predict vehicle motion, aligning the learning cycle closely with the physical dynamics observed on the race track. The strategic deployment of a local linear regressor further enhances the ATV model's accuracy.
3. Computational Efficiency: The localized data utilization profoundly reduces the computational time needed to solve the control problems. The results indicated that the finite time optimal control problem could be resolved in under 10ms on average, significantly outperforming previous implementations that required up to 90ms.

The research provides a framework that could have substantial implications for the future development of autonomous racing vehicles and potentially other high-speed dynamic applications. By demonstrating a scalable learning approach that optimizes computational resources while maximizing vehicle performance, the paper sets the groundwork for integrating such strategies into more sophisticated AI-driven vehicles.

Future Developments

Looking forward, this work could be expanded upon by integrating more complex models of tire dynamics and environmental interactions, which are intrinsic to high-performance racing scenarios. Additionally, extending the approach to more varied and intricate track configurations could yield insights into the broader applicability of the proposed LMPC framework.

Overall, this paper underscores the pivotal role of iterative learning and localized computational optimization within the rapidly evolving domain of autonomous vehicle control, marking a significant stride for real-time applications in challenging environments.