Roboticist argues better action representations may beat scaling synthetic data
KyleMorgenstein · x · 2026-11-17
Robotics researcher Kyle Morgenstein discusses a class of action-representation methods (including his CoRL work and @RobobertoMM's): find a better representation first instead of relying on massive synthetic data generation. He poses an open question — how far can robot learning get by simply collecting more data, and where does that break down? Beyond task success, he argues this line of work matters for safety and sample efficiency: images contain far more information than policies exploit, and efficiently extracting the subspace needed for control can change policy behavior.
More from Embodied
- Robotics researcher: edge cases are literally the entire problem for robot deployment — chris_j_paxton · 2026-11-17
- Better representations first: where does data scaling break down in robot learning? — KyleMorgenstein · 2026-11-17
- Policies use only a sliver of image information — extracting it matters for safety — KyleMorgenstein · 2026-11-17
- Enactra shows agentic robot simulation with controllable worlds and per-frame annotations — Lianhuiq · 2026-11-17
- South Korea's robot density hits 1,066 per 10k workers, 7x the world average — randal_olson · 2026-11-17
- Verne Robotics shows open-weight world action model V0.5 beating π0.5 on RoboTwin 2.0 with ~2hrs data — chris_j_paxton · 2026-11-17