Zhiyuan Open-Sources First Real-World Physical Agent Dataset

The significance here is less about dataset size than about the design shift: moving from manipulation difficulty to decision difficulty reframes what a physical agent must learn, folding goal understanding and process memory into the same loop as action execution. That matters most for teams training embodied models, who gain a real-world, reasoning-oriented alternative to purely motor-skill benchmarks. The pairing of high-level Q&A with low-level trajectories is the piece to watch, since it determines whether decision-level training actually transfers to robot execution. Whether nine scenarios prove broad enough for generalization is the open question.