Data-Driven Simulation for Training High-Performance Polices

Grzegorz Czechmanowski, Jan Węgrzynowski, Piotr Kicki, Krzysztof Walas

Keywords: Data-Driven Simulation, Reinforcement Learning, Real to Sim to Real

TL;DR: We learn a simulation model from real data, study what enables sim-to-real transfer, and deploy a racing policy zero-shot on F1TENTH - outperforming MPC and an expert human.

Abstract:

Autonomous racing is challenging control task - minimizing lap time requires operating at the limits of vehicle handling while strictly avoiding crashes. This makes simulation a compelling setting for learning racing policies, where high-risk exploration is safe and scalable, provided the simulator is sufficiently accurate to enable transfer to real-world.
We therefore investigate the critical components of a real→sim→real pipeline: learning dynamics models from real-world data, training the policy in simulation, and transferring it zero-shot to a real F1TENTH car. We demonstrate the effectiveness of this pipeline by outperforming both a nonlinear Model Predictive Control (MPC) baseline and an expert human driver.

Serve As Reviewer: ~Grzegorz_Czechmanowski1, ~Jan_Węgrzynowski1

Submission Number: 13