Research map / Reinforcement learning / Visual and representation RL
A Simulation Benchmark for Autonomous Racing with Large-Scale Human Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Assetto Corsa Gym, a high-fidelity racing simulation platform with a Gym/ROS interface, benchmarks SAC, TD-MPC2 and MPC on four tracks and three cars, and releases a 2.3M-step human driving dataset plus 64M SAC replay steps.
Problem
Autonomous racing research is limited by the high cost of race cars and by the limited physics accuracy of open-source simulators, and no open racing platform offers human driving data at the limits of handling.
Approach
A Gym-compliant and ROS2-compatible interface on top of the Assetto Corsa simulator (via its plugin API and a virtual joystick) that enables RL and MPC benchmarking and recording of human demonstrations, with human-driving data used to bootstrap offline-to-online RL (SAC fD, TD-MPC2 fD) and multi-track pretraining/finetuning.
Key results
- SAC fD achieved 104.08 s on Monza with F317, beating the best human (104.31 s) and SAC from scratch (104.41 s).
- With the GT3 car, the best human lap (108.38 s on Barcelona) remained faster than SAC fD (108.96 s), and SAC trained from scratch took about 4 days to match the best human.
- TD-MPC2 finetuned from multi-track pretraining completed laps on unseen tracks after a few episodes, while SAC fD and TD-MPC2 fD needed hundreds of episodes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| initialize SAC with human driving demonstrations | SAC · F317, across tracks | lap time: 97.67 -> 97.54 (BRN); 83.51 -> 83.61 (RBR); 104.41 -> 104.08 (MNZ) | Mixed |
| initialize SAC with human driving demonstrations | SAC · GT3, across tracks | lap time: 109.51 -> 108.96 (BRN); 93.11 -> 92.02 (RBR); 112.69 -> 111.64 (MNZ) | Helped |
| pretrain TD-MPC2 on other tracks, then finetune on the target track | TD-MPC2 · MNZ and AT target tracks | convergence speed | Helped |
| pretrain TD-MPC2 on other tracks, then finetune on the target track | TD-MPC2 · MNZ and AT target tracks | crashes during training | Helped |
| use TD-MPC2 fD rather than SAC fD | TD-MPC2 fD and SAC fD · target tracks during training | crashes during training | Helped |
| use expert rather than beginner human demonstrations | SAC · BRN / F317 | lap time and convergence | Mixed |
| remove the reference path (ablation) | SAC trained from scratch · F317 and GT3 | lap completion: 97.67 -> Fail (F317); 109.51 -> Fail (GT3) | Hurt |
| remove the reference path (ablation) | SAC fD · F317 and GT3 | lap time: 97.67 -> 99.80 (F317); 109.51 -> 110.98 (GT3) | Hurt |
Surprising findings
- SAC trained from scratch fails to complete a lap without the reference path, but SAC fD succeeds, suggesting human data helps exploration.
- Beginner demonstrations speed up early RL training but lead to slower final lap times, possibly due to bias toward human behavior.
Limitations
- Assetto Corsa runs only in real time, limiting faster-than-real-time training and requiring parallel workers.
- No image observations were explored; policies rely on range finder and telemetry inputs.
- RL policies are sometimes outperformed by expert humans, and automatic gearbox limits performance.
Experimental setup
Datasets: human driving data collected in Assetto Corsa SAC replay buffer data
Benchmarks: Assetto Corsa lap time on Indianapolis, Barcelona, Red Bull Ring, Monza
Metrics: best lap time lap time vs. episodes cumulative crashes (did not finish)
Baselines: Built-in AI Model Predictive Control SAC TD-MPC2
Builds on: Soft Actor-Critic TD-MPC2 Assetto Corsa
Closest papers by meaning
- NAVSIM: Data-Driven Non-Reactive Autonomous Vehicle Simulation and Benchmarking (NeurIPS 2024 · similarity 0.85)
- Beyond Self-Play and Scale: A Behavior Benchmark for Generalization in Autonomous Driving (NeurIPS 2026 · similarity 0.85)
- Robust Autonomy Emerges from Self-Play (ICML 2025 · similarity 0.85)
- ReSim: Reliable World Simulation for Autonomous Driving (NeurIPS 2025 · similarity 0.84)
- IntersectionZoo: Eco-driving for Benchmarking Multi-Agent Contextual Reinforcement Learning (ICLR 2025 · similarity 0.84)
- Bench2Drive: Towards Multi-Ability Benchmarking of Closed-Loop End-To-End Autonomous Driving (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.