Research

Driving policies trained end to end, combining reinforcement learning with diffusion and flow-matching policy classes.

Constraining trajectory optimization at the endpoint to steer a driving foundation model, evaluated closed-loop in a simulator reconstructed from real-world data rather than on logged trajectories.

A multiscale, GPU-accelerated simulation framework for high-fidelity digital twins of chemistry labs — robotic manipulation, powder and liquid dynamics, heat transfer and reaction kinetics, driven by a modular semantics engine.

Adapted open geospatial foundation models to downstream Earth-observation tasks for Carbon Mapper and Climate TRACE, and built the evaluation around them.