NVIDIA ICRA'26 keynote: human data is the most scalable source for robot foundation models
yukez · x · 2026-09-29
An NVIDIA researcher shared the video and slides of their ICRA'26 keynote from three months ago, with three key messages:
- Building human-like machines, a centuries-old dream, is rapidly becoming reality.
- Human data is by far the most scalable source of training data for robot foundation models.
- World models act as a "data sponge" absorbing massive multimodal data for policy generalization.
The talk synthesizes a year of work — GR00T, EgoScale, SONIC, WAM — and shows a demo of a model learning to assemble the YCB airplane with zero teleop data. More progress is teased.
More from Embodied
- VSArena V1 ships: evaluate embodied/VLA agents without physical robots — NovaCoding · 2026-09-29
- French startup Wandercraft wins FDA clearance for hands-free home exoskeleton — LexiLove · 2026-09-29
- Samsung software update bricks AI smart refrigerators in South Korea, spoiling food — Polymarket · 2026-09-29
- Hand-sized 6-axis robot arm hits sub-micrometer precision for semiconductor work — LexiLove · 2026-09-29
- Aero UMI and Aero Hand launch at IROS 2026: capture rig plus high-DOF dexterous hand as one system — LexiLove · 2026-09-29
- Maker shows rough V0.1 prototype held together with rubber bands, video coming soon — Aryvyo · 2026-09-29