We scale reinforcement learning and connect it to embodied control — architectural and normalization advances that let deep RL scale with compute (SimBa), and vision-language-action models that guide robot manipulation through action-coherence and metric 3D trajectory prediction.
Selected Publications — 5 most representative of 12
SimBa injects a simplicity bias via input normalization and residual blocks, letting deep RL scale up parameters for better sample efficiency
SimbaV2 stabilizes RL optimization using hyperspherical weight normalization and reward-scaled value estimation, scaling with compute to state-of-the-art continuous control
