Nature Will Releases Natura-1: 50B MoE Model Enables Zero-Shot Generalization and Dexterous Manipulation

Natura-1, released by Nature Will, is a 50B-parameter Mixture-of-Experts (MoE) model that takes language and vision as input and drives a 20-DoF dexterous hand within a 100-millisecond control loop. It targets two capabilities: dexterous manipulation relying on the dexterous hand to achieve multi-finger coordination, object pose adjustment and complex contact, and zero-shot instruction following—directly completing operations based on natural language instructions when facing unseen scenarios and tasks.

These two goals have often sat at opposite ends of past approaches: either general policies aimed at grippers, or dexterous manipulation systems built for a single task. Natura-1 attempts to cover both with the same model. The team told Machine Heart that choosing human data and choosing a dexterous hand were the same decision—many tools and manipulation methods in home and industrial scenarios are designed around the human hand, and using a dexterous hand preserves richer hand-motion information, reducing the information loss incurred when compressing human hand movements into gripper open-close.

Training proceeds in two steps: first pretraining on 1 million hours of human experience data, then one round of efficient post-training on limited robot data to adapt the world knowledge learned in pretraining to dexterous manipulation. The team says it followed the approach from large language model research of scaling model size and data size together, first using smaller models and data to quickly validate the architecture, and gradually scaling up after confirming the direction.

The team released six demonstrations at original speed, in one take, with no human intervention: wrapping wontons, brewing tea, making a yogurt bowl, packing a package, applying a phone screen protector, and writing calligraphy with a brush—the longest running over 3 minutes. Each contains multiple subtasks, and the robot must continuously handle liquids, granules, soft tools and flexible components within the same workflow.

Take applying the phone screen protector: Natura-1 first wipes the screen, then fits the pre-loaded alignment frame with the protector onto the phone, aligning and pressing it down with both hands; then one hand holds the phone and alignment frame against the fixture while the other hand smoothly pulls out the release film, and finally it presses along the edges and corners with its fingertips to push out air bubbles. When packing a package, it must switch grasping methods among raffia, plush toys and a notebook, with one hand holding down the box lid that tends to spring back while the other hand pushes the tape dispenser to seal the box, and finally applies the label and scans the barcode.

These actions, while being executed, are changing the shape of objects and the contact conditions, and the model must continuously correct itself based on real-time feedback rather than following a fixed trajectory to completion. But it must be emphasized that these six demonstrations showcase manipulation depth after task fine-tuning; whether the capability can be brought to tasks it has not practiced depends on the zero-shot tests.