DeepMind's Generalization by Construction: inference-driven robot generalization
du_yilun · x · 2026-09-09
Du Yilun published a research perspective, "Generalization by Construction," addressing how robots can generalize beyond collecting more demonstrations. The core idea: with learned world models, a robot can plan future actions and goals via inference before acting, solving tasks it was never explicitly trained on—combining learning with inference for flexible generalization.
Related event: DeepMind Researcher Proposes Reasoning-Driven Robot Generalization(3 posts)→
More from Embodied
- Drone footage shows Tesla's Optimus factory rising fast, designed for up to 10M robots a year — XFreeze · 2026-09-09
- NumeriaPlus: 300-hour egocentric activity dataset with dense ground truth for world models — ducha_aiki · 2026-09-09
- Light REACT trains robots to adapt on the fly to hardware damage — chris_j_paxton · 2026-09-09
- Robotics researcher: foundation models are the future, but the problem space still matters — chris_j_paxton · 2026-09-09
- Industry's embodied CV data dwarfs academia's, chart shows log-scale gap — ducha_aiki · 2026-09-09
- johnowhitaker: shelf-stocking, burger-flipping humanoids by September 2031 — johnowhitaker · 2026-09-09