A Simulation Benchmark for Autonomous Racing with Large-Scale Human Data

TL;DR

The paper introduces a high-fidelity, Assetto Corsa–based benchmark for autonomous racing that enables reproducible evaluation of RL and MPC algorithms and collects a large-scale human driving dataset. It formalizes a Gym-like environment, implements SAC and TD-MPC2 alongside MPC and a built-in AI, and demonstrates that human demonstrations and cross-track pretraining significantly boost data efficiency and generalization. Key findings show that SAC with expert data can rival or exceed top human performance on certain tracks, while pretraining across tracks enables rapid adaptation to unseen tracks with fewer crashes. The work provides open-source code, datasets, and benchmarks to accelerate research in fast, safe, and generalizable autonomous racing systems.

Abstract

Despite the availability of international prize-money competitions, scaled vehicles, and simulation environments, research on autonomous racing and the control of sports cars operating close to the limit of handling has been limited by the high costs of vehicle acquisition and management, as well as the limited physics accuracy of open-source simulators. In this paper, we propose a racing simulation platform based on the simulator Assetto Corsa to test, validate, and benchmark autonomous driving algorithms, including reinforcement learning (RL) and classical Model Predictive Control (MPC), in realistic and challenging scenarios. Our contributions include the development of this simulation platform, several state-of-the-art algorithms tailored to the racing environment, and a comprehensive dataset collected from human drivers. Additionally, we evaluate algorithms in the offline RL setting. All the necessary code (including environment and benchmarks), working examples, datasets, and videos are publicly released and can be found at: https://assetto-corsa-gym.github.io

Figures (15)

• Figure 1: Overview. We propose a high-fidelity racing simulation platform based on Assetto Corsa that enables reproducible algorithm benchmarking, as well as data collection with human drivers.
• Figure 2: Assetto Corsa. A GT3 (top) and a F317 (bottom) car each turning a corner in the Assetto Corsa simulator. The simulator features a total of 178 official cars and 19 laser-scanned tracks, in addition to custom content created by the community. We develop a platform for interfacing with the simulator that can be used with both RL and MPC methods, as well as human drivers.
• Figure 3: Our proposed platform for autonomous racing. We provide interfaces (gray) that (1) connect a simulator (Assetto Corsa) to autonomous racing methods, and (2) allow for human data collection. Interfaces receive track information and state, and execute actions in the simulator. Datasets (purple) are collected using an ACTI (Assetto Corsa Telemetry Interface) tool.
• Figure 4: Cars and tracks. We consider a total of 4 different tracks (left), as well as 3 distinct cars (right). We collect and open-source human driving data for all tracks and cars considered.
• Figure 5: Mean delta lap time vs. best human. Across all 4 tracks. SAC fD performs better than the best human with F317 but worse with GT3.