Radar / AI robotics / Natura-1
Natura-1: 50B robot foundation model trained on 1M hours of human video
Nature Will's 50B-parameter mixture-of-experts vision-language-action model pretrained on 1M+ hours of egocentric human video, driving 20-DoF dexterous hands on dual-arm robots at a 100ms observation-to-action loop.
Why it matters
Instead of per-task robot policies, it proposes one reusable manipulation policy: pretrain on scaled human experience (dense 3D hand poses recovered from video), then adapt with limited robot data.
What you could build with it
A robotics lab could pretrain a shared vision-language-action policy on annotated human video the way Natura-1 does, then fine-tune it with a small robot dataset, getting one policy that handles unfamiliar manipulation tasks instead of a library of single-task policies.
Does it hold up?
Too early to judge — no public weights, API, code, or independent evaluation; evidence is limited to company demos and first-party benchmark claims.
Built with Natura-1
- Natura-1 Brings a 50B Model to Dexterous Robot Handsarticle · Critical walkthrough of the claims: shared-weights policy thesis, the human-video pretraining pipeline, why six long-horizon demos used extra tuning, and what the 100ms loop does and does not prove.
Learn more
First spotted on hn: source.
More AI robotics
FLUX 3 ActionBlack Forest Labs launched FLUX 3 Action on Sept 24: a 7B open-weight world-action robotics model that takes camera…robotics · JEV 0.72RoboParty RP1 (ROBOTO 01)Full-stack open-source humanoid robot unveiled at IROS 2026 with Romomo actuators, PartyOS software, 160 N·m peak…robotics · JEV 0.67Boston Dynamics fits Atlas humanoid with a factory-ready four-fingered handAtlas's new hand has 13 degrees of freedom (nearly double the prior three-fingered generation) with pressure sensors…robotics · JEV 0.66NVIDIA Isaac ROS 5.0: agentic AI workflows for roboticsIsaac ROS 5.0 moves to ROS 2 Lyrical/Ubuntu 24.04 and adds agent-ready skills - reusable workflows like…robotics · JEV 0.59Maxwell: embodied AI model tops Meta-World physical-task benchmarkResearchers at the Chinese Academy of Sciences' Institute of Artificial Intelligence for Industries reported Maxwell…robotics · JEV 0.54Light-O1 (Light Origins): Embodied Foundation ModelGeneral-purpose embodied foundation model that learns a reusable human action prior from structured human actions…robotics · JEV 0.43
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.