Embodied AI's 'Pre-GPT Era' Dilemma: Lack of Foundation Models and Broken Data Flywheels

If I had to sum up embodied AI at WAIC 2026 in one word, it would be "disconnected." On the exhibition floor, robots fold clothes, brew coffee, and chop veggies, looking like they've nailed every task. But step off the booth and chat with people actually working on deployment, and you hear almost nothing but "don't rush," "it's too early," or "that's not how we do it." Yao Yuan from ModelBest drew a parallel, comparing today's embodied AI to NLP in the "pre-GPT era"—training separate models for each task, lacking a universal foundation model, and missing out on data flywheels. I deeply agree with this assessment, especially since I've been reading several papers on robot foundation models recently. The analogy is not just apt; it hits right at the industry's core bottleneck.