Nature Will unveils Natura-1: 50B MoE robot model trained on 1M+ hours of human video
stingning · x · 2026-10-13
Nature Will introduced Natura-1, its first general foundation model for the physical world: a 50-billion-parameter MoE pretrained on over 1 million hours of egocentric human experience. It takes language and vision input and drives 20-DoF dexterous hands in a 100ms control loop. The recipe: pretrain world knowledge on massive human data, then efficiently post-train on limited robot data, yielding zero-shot natural-language instruction following that generalizes to unseen scenarios. The technical blog covers real-world demos, scaling results, training data, architecture, and infrastructure.
More from Embodied
- Scoble: A 2018 Model 3 on HW3 Now Drives Nearly Like a Cybercab After FSD 14.3 Update — Scobleizer · 2026-10-13
- WUJI's dexterous robot hand packs 20 active DoF with high-frequency control — Olivier__OG · 2026-10-13
- Robotics startup Shita raises angel round backed by Huawei's Hubble for sub-1mm photonic-crystal e-skin — 量子位 · 2026-10-13
- Humanoid hand wraps wontons as Nature Will releases Natura-1, a foundation model trained on 1M hours of human data — CyberRobooo · 2026-10-13
- Natura-1 runs multi-minute autonomous dexterous hand workflows with a single policy — stingning · 2026-10-13
- Natura-1 demo: a single-take 30-minute recording of everyday human-robot interaction — stingning · 2026-10-13