Nature Will unveils Natura-1: 50B-parameter MoE robot foundation model trained on 1M hours of egocentric human video
rohanpaul_ai · x · 2026-10-14
Nature Will introduced Natura-1, its first general foundation model for the physical world: a 50B-parameter MoE pretrained on 1M hours of egocentric human video, then post-trained on a limited amount of robot data for dexterous manipulation.
Key points:
- Follows complex natural-language instructions zero-shot, reasoning over spatial constraints, multi-object relations, and negative conditions
- One model handles three things robots usually fail at together: fine two-handed control after grasp, new tasks without extra training, and long sequences requiring memory and error recovery
- The bet: human footage is a faster path to scale than slow robot demonstrations, provided skills survive the human-to-robot hand transfer
More from Embodied
- Dyna-2 Deep Dive: World-Action Models Trained on 1M Hours of Human Video, 99.4% 24-Hour Autonomy — Latent Space · 2026-10-14
- 15 open-source ESP32 AI gadget projects plus a step-by-step Meta Muse SDK flashing tutorial — alexcovo_eth · 2026-10-14
- Collaborative proprioception: Auki gives a phone and robot a shared sense of space — broodsugar · 2026-10-14
- Apple just released a new smart-home command center: should you buy it? — Kiteaton · 2026-10-14
- Physical Intelligence launches Pi API beta for fine-tuning π robotic foundation models — chelseabfinn · 2026-10-14
- VPP2: video-prediction robot policy beats Cosmos3-64B by 11%, tops RoboDojo — lucy_x_shi · 2026-10-14