Ryo Hanai

Stochastic action tokenization unlocks significant performance gains in robot policy learning, especially under data scarcity, by leveraging representational redundancy.

Achieving a 97% success rate in real-world non-prehensile throwing, this approach redefines the limits of robotic object transport beyond traditional grasping techniques.

The embodiment gap reveals that even advanced robot foundation models require substantial adaptation to function on specific robotic platforms, challenging the assumption that scaling alone suffices for generalization.

Immediate feedback-driven corrections can enhance robotic manipulation performance without the overhead of retraining entire policies.

Correctable visual attention can drastically enhance robot manipulation performance, especially in unpredictable environments.