RSS 2026 shows embodied AI is still debating VLA, world models, and data curation
量子位 · wechat · 2026-07-25
RSS 2026 surfaced a more nuanced embodied-AI consensus: VLA and world models are complementary
At RSS 2026, researchers and founders repeatedly pushed back on the idea that “world models will replace VLA.” A common view was that the field is still far from maturity, and that the real problem is not choosing one camp but understanding what a world model should predict and how those predictions help robot action.
- Speakers from Physical Intelligence and other labs argued that VLA and world models are not mutually exclusive.
- For many tasks, language still matters for task decomposition and instruction following.
- Video prediction alone may produce plausible futures, but without action-related signals such as hand size, object size, and control constraints, it may not translate into robust manipulation.
- Some researchers proposed separating dynamics prediction from utility evaluation: one model generates futures, another judges whether those futures help the task.
The article also emphasizes that robotics is increasingly a data-curation problem, not just a scaling problem:
- Egocentric human data helps, but should not be bought blindly; curation, filtering, and mixture design matter more than raw volume.
- Useful robot learning often comes from a combination of robot teleoperation data, task-aligned human demonstrations, and free-form human video.
- Karen Liu’s keynote framed the field as “Data-poor, Model-rich”: existing simulation, graphics, planning, and control models can be turned into teachers that generate training data.
- Her team’s VLK pipeline reportedly generated 48,000 visual-language-full-body trajectories in a reconstructed indoor environment and transferred to a real Unitree G1 for navigation and object transport.
The piece closes by noting that industry discussions in China are also shifting toward hardware advantage, pricing, ROI, and local deployment scenarios such as mining and inspection.
More from Embodied
- Robot goes to the fridge and fetches a beer — Darpinian · 2026-07-27
- Researchers show digital circuits can be replicated with knitted fabric — mtizard · 2026-07-27
- Local Qwen models power a robot that tests 78 smartphones’ battery life — gappyvalley · 2026-07-27
- TechCrunch says brain-wave signals could be the next unlock for physical AI training — TechCrunch AI · 2026-07-27
- YC talk says robots need better pretraining, memory, and compositional behaviors — ycombinator · 2026-07-27
- China’s drone-license boom cools as training centers shrink and professional demand takes over — 创业邦 · 2026-07-27