World models, neural simulation, and learning in interactive environments.
Investigated transfer in long-horizon model-based reinforcement learning by training PPO across 2.5 billion steps inside a 0.66M-parameter world model. Using an 8-GPU A40 fleet and Dyna anchoring, policies achieved a 70.2% holdout win rate.
Can a learned world model replace the simulator itself, rather than just predict its next frame? Compact Dreamer/Genie-style models (1.6M and 3.8M parameters) run CARLA driving and a browser flight game fully autoregressively at 30 fps on a laptop CPU.
Studied closed-loop driving limits from offline data with a 1.36M-parameter multimodal policy using temporal RGB, LiDAR, and telemetry. Trained on 3.3 hours of expert demonstrations, it achieves hours of zero-intervention driving, cross-town transfer, and robust error recovery.
Also on Google Scholar and arXiv.