Speakers

Real-world robotic reinforcement learning is fundamentally constrained by expensive and risky data collection, makingsample efficiency, robustness, and reasoning abilityessential for practical deployment. However, existing approaches often suffer from inefficient exploration, limited state coverage, and unstable policy updates when interacting with complex physical environments. To address these challenges, we develop a unified research line that improves real-world learning throughinformative sample selection, diversity-aware exploration, data augmentation, reasoning-enhanced decision making, and structured policy optimization. Concretely, our works include: (1) entropy-guided sample prioritization and diversity-driven exploration to increase effective data utilization, (2) exploratory data augmentation for manipulation robustness, (3) online reasoning mechanisms for Vision-Language-Action models, and (4) explicit manifold alignment to stabilize online policy learning. These efforts collectively advance practical robotic learning systems that learn faster, generalize better, and operate more reliably in real-world scenarios.

Robotic generalist policies are trained on large-scale, diverse datasets, thus internalizing many useful behaviors. We posit that semantic knowledge can elicit appropriate behaviors to solve new tasks, even if directly commanding the generalist policy fails to do so. We present two mechanisms for this: First, by training Steerable Policies: vision-language-action models (VLAs) that can follow a much wider range of language commands, spanning a range of abstractions -- from subtask-level plans to pixel coordinates. In accepting this spectrum of "steering commands," the policy more readily makes use of vision-language models' skills for planning and in-context reasoning. It also enables Semantic Action Reinforcement Learning (SARL) -- wherein we learn which prompts and abstractions reliably cause the VLA to solve the task, thus treating generalist policy prompts as an action space. Alternatively, we can use semantic reasoning to pick out the "reasonable" skills from the generalist's wide action prior. Specifically, we propose Flow Reversal Steering (FRS): a method for refining rough sketches of good robot behaviors into similar precise actions by passing them through generalist flow policies in reverse. We show how, while VLM reasoners struggle with directly outputting low-level actions, FRS allows them to focus on making high-level semantic inferences, which the generalist "projects" into good actions. In turn, FRS can bootstrap RL by exploring semantically-reasonable interactions, thereby improving and accelerating learning in tasks that base policies completely fail at. With both Steerable Policies and FRS, we hope to show a new paradigm for generalist reinforcement learning, wherein policy improvement is guided with both rewards and reasoned semantic knowledge.