On Oct 13 Nature Will introduced Natura-1, a 50B-parameter MoE pretrained on 1M hours of egocentric human data and post-trained on limited robot data. It drives 20-DoF dexterous hands in a 100 ms loop and follows natural-language instructions zero-shot. On the lab's own benchmark it scores 53.4 (Base) / 59.1 (Agentic); π0.5 and GPT-6-Astra score 13.5 and 16.7 on a gripper platform. No weights or API are released.
Key Takeaways
- ✓50B MoE jointly modeling language, images, video and action
- ✓1M hours pretraining: 61.7M episodes, 317,842 tasks, 14,286 scenes
- ✓Scaling 44K→1M hours cuts real-robot action-prediction MSE by 67.7% (no robot fine-tuning)
- ✓Benchmark 53.4/59.1 vs π0.5 13.5 and GPT-6-Astra 16.7 (baselines on grippers)
- ✓100 ms observation-to-action loop; >50% MFU in training
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Why it matters
Robot foundation models usually cover either general policies for simple grippers or task-specific systems for dexterous hands. High-DoF hands leave little room for error at contact, and robot data can only be collected where robots already work. Nature Will, a lab founded in April 2026 and backed by ZhenFund, launched Natura-1 on October 13, 2026 with a different bet: learn from human first-person experience at scale.
How it works
Natura-1 is one 50B mixture-of-experts transformer trained jointly on language, images, video and action. Each episode is modeled autoregressively as a single sequence; modality-specific bias lives only in token layout and attention causality. Video and action are treated as two views of one physical state and trained together with flow matching and joint denoising, so abundant video adds supervision to scarce action labels. Pretraining uses 1M hours of egocentric data (61,702,471 episodes, 317,842 tasks, 14,286 scenes), and the lab reports a 41% lower MPJPE for its hand-motion labels than open-source methods. A limited amount of robot data then adapts the model to dual-arm robots with 20-DoF hands. Training holds over 50% MFU on hundreds of GPUs, and a Rust orchestrator plus fused CUDA kernels keep a strict 100 ms observation-to-action loop.
Numbers (lab-reported)
- 12-dimension dexterous benchmark: Natura-1 53.4 (Base) and 59.1 (Agentic). π0.5 and GPT-6-Astra (used as a coding policy) scored nothing on the dexterous-hand rig; on a gripper platform they reached 13.5 and 16.7. The hardware differs, so this is not a like-for-like comparison.
- Scaling egocentric data from 44K to 1M hours cut real-robot action-prediction MSE by 67.7% without robot fine-tuning.
- Six multi-minute workflows (wontons, packing and taping a box, tea service, calligraphy, screen protector, yogurt bowl) use extra task-specific tuning; other results share one set of weights.
What developers can do
There are no open weights, API or SDK yet; the technical blog is the primary source. Embodied-AI teams can reuse the recipe: human-video pretraining with light robot post-training, whole-trajectory sequence modeling, joint video-action denoising, and taxonomy-driven data collection that fills rare tasks. Treat the benchmark as vendor-built until third parties reproduce it.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.