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

Adrian Remonda1,2,4,  Nicklas Hansen1,  Ayoub Raji3,  Nicola Musiu3,
Marko Bertogna3,  Eduardo E. Veas2,  Xiaolong Wang1

1UC San Diego,  2TU-Graz,  3Unimore,  4Know-Center

Render of Assetto Corsa
FPV of Assetto Corsa
Our advanced simulator
Our mid-range simulator

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.

AssettoCorsaGym. Racing simulation platform using Assetto Corsa for autonomous driving. Includes a human driver dataset, and a comparison of different algorithms.

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, and datasets are publicly released and can be found here.

Videos

SAC from scratch vs SAC from demonstrations

SAC from Scratch vs. Using Human Demonstrations. Increased Jitter in training SAC from Scratch (left), while human demonstrations make the controls smoother (right).

Recovering Manuevers

Recovering maneuver. Delays are added artificially to force the car to lose control and recover it.

Human data collection

Supporting Open-Source Science

Tracks and cars. The four tracks and three cars used in our dataset. The tracks include Indianapolis (IND), an easy oval track; Barcelona (BRN), featuring 14 distinct corners; Austria (RBR), a balanced track with technical turns and high-speed straights; and Monza (MNZ), the most challenging track with high-speed sections and complex chicanes. The cars are the Mazda Miata NA (Miata) with a top speed of 197 km/h, the Dallara F317 (F317) with a top speed of 250 km/h, and the BMW Z4 GT3 (GT3) with a top speed of 280 km/h. This diverse array ensures a comprehensive dataset for evaluating driving algorithms.

We open-source a 64M-step dataset with 2.3M steps from human drivers and the remaining steps from Soft Actor-Critic (SAC) policies. Data was collected at UC San Diego and Graz University of Technology, involving 15 drivers completing at least five laps per track and car. Participants included a professional e-sports driver, four experts, five casual drivers, and five beginners. Data can be downloaded manually using the links or via this script.

Car Track Human Data SAC Data Download
Stints Laps Steps Steps
F317 BRN 70 247 612,557 10M Download
F317 MNZ 19 117 288,582 10M Download
F317 RBR 24 142 295,679 10M Download
F317 IND 1 4 4,605 4M Download
GT3 BRN 37 181 501,206 10M Download
GT3 RBR 15 102 218,722 10M Download
GT3 MNZ 13 85 221,123 10M Download
Miata BRN 5 27 99,145 10M Download
Miata MNZ 2 10 38,395 - Download
Miata RBR 3 12 32,971 - Download
Total 189 927 2,312,985 64M

A Simulation Platform for Autonomous Racing

Overview. 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.

Citation

If you find our work useful, please consider citing the paper as follows:

@misc{remonda2024simulation, title={A Simulation Benchmark for Autonomous Racing with Large-Scale Human Data}, author={Adrian Remonda and Nicklas Hansen and Ayoub Raji and Nicola Musiu and Marko Bertogna and Eduardo E. Veas and Xiaolong Wang}, booktitle={arxiv}, year={2024} }