The authors pretrained a learned world model, which predicts how the vehicle and environment evolve, using heterogeneous trajectories from low- and high-fidelity simulators.
Simulation priors speed adaptation of model-based autonomous racing controllers
PIA pretrained a world model on mixed-fidelity simulated driving data, then corrected simulation mismatch through limited interaction with new tracks, surfaces and a full-size race car.
Big Tech
Adrian Remonda · Jiajun Xi · Nicklas Hansen · Marcus Greiff · Thomas Lew · John Talbot · +2 more
Toyota Research Institute · University of California San Diego (UCSD)
Research Digest··3 min read
Remonda et al.
§