同步下 WAIC 逛完一天的感受，最大的共识是具身的数据还远未收敛。

TL;DR - A WAIC forum highlighted data—not model architecture—as the main bottleneck to scalable embodied AI. Industry leaders are exploring real-robot, egocentric human, video, wearable, and deployment data to build generalizable physical intelligence.

• Key barriers are scarce interaction data, missing cross-task/robot representations, and costly real-world feedback loops.
• Physical Intelligence and Dyna emphasized learning from failures and recycling deployment data into training; Dyna reported 99.4% success across 850+ napkin-folding trials after RL post-training.
• Speakers advocated combining robot, human egocentric, UMI, video, and internet data to improve diversity and transfer across embodiments.
• Estimates for an embodied-AI “ChatGPT moment” ranged from under two to five years, largely depending on progress in data collection and closed-loop learning.