01 / EXECUTIONRobotic Agent Systems
How can a robot turn an open-ended goal into reliable, verifiable action?
We connect task-level reasoning with VLA skills, studying how agents plan, compose behaviors, and check their effects in the physical world.
Our focus is the transition between skills: tracking state, detecting failed preconditions, and recovering when execution diverges from the plan.
Long-horizon planningSkill compositionVerification & recovery
02 / PRE-TRAININGEmbodied Pre-training
What data and representations make physical skills transfer?
We study vision-language-action and world-action models trained on UMI, robot, and egocentric experience, with the aim of generalizing across tasks and embodiments.
Data quality is part of the learning problem. We investigate how capture fidelity, temporal alignment, coverage, and data mixtures affect real-robot performance as training scales.
VLA & world-action modelsHigh-fidelity dataCross-embodiment transfer
03 / POST-TRAININGContinual Learning & RL
How can experience improve both a robot’s skills and its ability to learn?
We are building Simple Evolver to learn from success, failure, and human feedback through real-robot reinforcement learning. World and value models help us study action outcomes, evaluate behavior, and guide policy updates.
We explore meta-learning, online adaptation, and recursive self-improvement (RSI): carrying experience into new tasks while retaining earlier capabilities, with gains tested through repeated deployment.
Real-robot RLWorld & value modelsMeta-learning & RSI