Yutaka Matsuo

Robots can now intelligently choose camera angles to answer questions by predicting the utility of unseen views, bypassing costly real-time rendering.

Frozen 4B image generators can outperform proprietary models at multi-subject composition simply by letting an autonomous coding agent iteratively rewrite their Python orchestration harnesses.

NavWAM turns visual foresight into executable robot actions, outperforming traditional planning methods in real-world navigation scenarios.