[Paper Review] Learn-to-Race Challenge 2022: Benchmarking Safe Learning and Cross-domain Generalisation in Autonomous Racing

Jonathan Francis, Bingqing Chen+11 more|arXiv (Cornell University)|May 5, 2022

Adversarial Robustness in Machine Learning6 citations

TL;DR

This paper presents the Learn-to-Race 2022 virtual challenge, a benchmarking platform using the L2R Task 2.0 simulation framework to evaluate reinforcement learning agents in autonomous racing with a focus on safe learning and cross-domain generalization. The challenge evaluated 733 model submissions from 46 teams across 88+ institutions, demonstrating that policy distillation and distribution-aware imitation learning significantly improved robustness and transfer performance on unseen racetracks under safety constraints.

ABSTRACT

We present the results of our autonomous racing virtual challenge, based on the newly-released Learn-to-Race (L2R) simulation framework, which seeks to encourage interdisciplinary research in autonomous driving and to help advance the state of the art on a realistic benchmark. Analogous to racing being used to test cutting-edge vehicles, we envision autonomous racing to serve as a particularly challenging proving ground for autonomous agents as: (i) they need to make sub-second, safety-critical decisions in a complex, fast-changing environment; and (ii) both perception and control must be robust to distribution shifts, novel road features, and unseen obstacles. Thus, the main goal of the challenge is to evaluate the joint safety, performance, and generalisation capabilities of reinforcement learning agents on multi-modal perception, through a two-stage process. In the first stage of the challenge, we evaluate an autonomous agent's ability to drive as fast as possible, while adhering to safety constraints. In the second stage, we additionally require the agent to adapt to an unseen racetrack through safe exploration. In this paper, we describe the new L2R Task 2.0 benchmark, with refined metrics and baseline approaches. We also provide an overview of deployment, evaluation, and rankings for the inaugural instance of the L2R Autonomous Racing Virtual Challenge (supported by Carnegie Mellon University, Arrival Ltd., AICrowd, Amazon Web Services, and Honda Research), which officially used the new L2R Task 2.0 benchmark and received over 20,100 views, 437 active participants, 46 teams, and 733 model submissions -- from 88+ unique institutions, in 58+ different countries. Finally, we release leaderboard results from the challenge and provide description of the two top-ranking approaches in cross-domain model transfer, across multiple sensor configurations and simulated races.

Motivation & Objective

• To establish a standardized benchmark for evaluating safe, high-performance, and generalizable reinforcement learning agents in autonomous racing.
• To address the challenge of robust perception and control under distribution shifts, novel obstacles, and unseen racetracks.
• To promote interdisciplinary research in autonomous driving by providing a high-fidelity, open-source simulation environment.
• To evaluate the joint capabilities of safety, performance, and generalization in multi-modal perception and control systems.
• To serve as a proving ground for sim-to-real transfer and safe RL through a two-stage competition format.

Proposed method

• The challenge uses the Learn-to-Race (L2R) Task 2.0 simulation framework, a high-fidelity, physics-based racing environment with multi-modal sensor inputs.
• A two-stage evaluation protocol was implemented: Stage 1 focused on high-speed racing under safety constraints; Stage 2 required adaptation to an unseen racetrack via safe exploration.
• Agents were trained using reinforcement learning with safety constraints enforced via cost-limit regressors, rules-based experts, or learnable safety critics.
• Baseline approaches included imitation learning with distribution-aware priors to improve robustness and reduce causal confusion during transfer.
• Model performance was evaluated using refined metrics including speed, safety compliance, and generalization to unseen tracks and sensor configurations.
• The platform supports integration with real-world vehicle stacks via ROS, enabling future sim-to-real transfer validation.

Experimental results

Research questions

• RQ1 How well can reinforcement learning agents generalize to unseen racetracks while maintaining safety and high performance?
• RQ2 What role do distribution-aware imitation learning priors play in improving robustness and reducing negative transfer in cross-domain racing?
• RQ3 How effective are different safety constraint mechanisms (e.g., rules-based, learnable critics) in maintaining sub-second, safety-critical decision-making?
• RQ4 To what extent can policy distillation improve generalization and performance in multi-modal perception-based autonomous racing?
• RQ5 Can the L2R Task 2.0 benchmark serve as a reliable proxy for sim-to-real transfer in autonomous racing?

Key findings

• The challenge attracted 46 teams from 88+ institutions across 58+ countries, with 733 model submissions and over 20,100 views, indicating strong community engagement.
• The two top-performing approaches leveraged policy distillation and distribution-aware imitation learning, achieving superior cross-domain transfer performance across multiple sensor configurations.
• Agents trained with distribution-aware priors demonstrated improved robustness to noise and better generalization on unseen racetracks compared to standard IL baselines.
• Safety constraints implemented via learnable critics and cost-limit regressors effectively prevented catastrophic failures during high-speed racing in dynamic environments.
• The L2R Task 2.0 benchmark successfully captured the complexity of real-world autonomous racing, including distribution shifts and novel obstacles, validating its use as a rigorous evaluation platform.
• The integration of the L2R framework with ROS enables future validation of algorithms on real-world autonomous racing platforms, such as those used in Roborace and Indy Autonomous Challenge.