Natura-1: Physical-World Foundation Model from Launch to Hands-On

Episode 1 · Nature Will Unveils Natura-1, a Foundation Model for the Physical World (2026-10-13, 9 posts)

Nature Will has released Natura-1, its first general-purpose foundation model for the physical world. Built on a 50-billion-parameter MoE architecture, it was pre-trained on roughly 1 million hours of real manipulation video captured from a human egocentric view, showing hands handling objects and scenes. The model takes language and vision as input, drives a 20-DoF dexterous hand via a 100ms control loop, and generalizes to unseen objects and tasks. A single model combines three capabilities: dexterous manipulation control, zero-shot natural-language instruction following, and long-horizon task execution.

Confirmed

• Released by robotics company Nature Will; the model, Natura-1, is its first general-purpose physical-world foundation model
• Scale and architecture: 50-billion-parameter MoE; pre-trained on 1 million hours of human egocentric experience video
• Takes language and vision input, 100ms control loop, drives a 20-DoF dexterous hand
• Claimed capabilities: zero-shot execution of complex natural-language instructions, generalization to unseen objects and tasks, long-horizon task execution
• Training approach: large-scale pre-training on massive human experience video to acquire world knowledge, followed by training for control tasks

Why it matters

• @lukasmziegler notes the long-horizon execution capability is what stands out: for long tasks, the state is determined by the model's own history, a key difficulty distinguishing it from short-horizon manipulation
• This validates the paradigm of pre-training robot foundation models on egocentric human video, unifying world knowledge and low-level control in a single model and offering a new scalable path toward general robotic manipulation

• Natura-1 debuts: 50B-param MoE world model trained on 1M hours of egocentric video — stingning · 2026-10-13
• Nature Will unveils Natura-1: 50B MoE robot model trained on 1M+ hours of human video — stingning · 2026-10-13
• Natura-1: a physical-world foundation model trained on ~1M hours of egocentric video — lukas_m_ziegler · 2026-10-13
• Natura-1: 50B-param MoE robot foundation model pretrained on 1M hours of egocentric human video — heyshrutimishra · 2026-10-13
• Nature Will launches Natura-1, a foundation model for the physical world trained on 1M hours of egocentric video — testingcatalog · 2026-10-13
• Nature Will launches Natura-1, a physical-world foundation model trained on 1M hours of egocentric video — testingcatalog · 2026-10-13
• Nature Will's Natura-1 gives robot hands zero-shot instruction following and dexterous control — mhdfaran · 2026-10-13
• 1X unveils robot foundation model trained on nearly 1M hours of first-person human data — LinusEkenstam · 2026-10-13
• Natura-1: 50B MoE foundation model drives dexterous hands zero-shot from plain language — stingning · 2026-10-13

Episode 2 · Natura-1 Robot Passes 30-Minute Zero-Shot One-Take Test (2026-10-13, 2 posts)

Nature's team and researcher Chris Paxton demoed Natura-1 running on the Wuji Hand 2 dexterous hand, completing a 30-minute unedited one-take test where the robot handled novel verbal instructions and unfamiliar environments with zero-shot generalization and self-correction on failure.

• Natura-1 zero-shot demo runs on Wuji Hand 2 in one uncut 30-min take — chris_j_paxton · 2026-10-13
• Natura-1 robot passes 30-minute one-take zero-shot test with spoken instructions — stingning · 2026-10-14