Preston Culbertson

Task progress in vision-language-action models can be read directly from their internal representations, even before task-specific training, revealing a surprising depth of interpretability.

REGRIND enables robots to master complex dexterous tasks from just one human demonstration, achieving impressive sim-to-real transfer without extensive retraining.

Unleash your legged robots: Sumo lets them dynamically manipulate objects far exceeding their weight and size limits, all without task-specific retraining.

Ditch slow iterative refinement: conditional flow-matching models can directly learn meaningful proposal distributions from noisy sampling-based MPC data, slashing planning time.