The Will to Act — Natura-1

TL;DR

Natura-1 is Nature Will’s first general physical-intelligence foundation model: a 50B-parameter mixture-of-experts transformer pretrained on 1 million hours of egocentric human experience (61.7M episodes, 317K tasks, 14.3K scenes), then post-trained on a limited amount of robot data to drive 20-DoF dexterous hands within a strict 100 ms control loop. It stakes the position that generalist dexterous manipulation falls out of human-video scale paid for by in-house data + training + inference infrastructure co-designed around a single MoE backbone that unifies visual world modeling and action generation, rather than teleop scaling, UMI collection-tool invariance, or VLM backbones with action heads. Reports zero-shot following of complex natural-language instructions (spatial constraints, multi-object relationships, negative conditions), autonomous multi-minute workflows across tool switches / deformable materials / compliant brushstrokes, and emergent recovery behaviors (sliding a card to create a graspable overhang, mid-air wrist rotation to steady a toppling stack) that were not in the robot demonstration data.

Key claims

Natura-1 is a single 50B-parameter mixture-of-experts transformer trained jointly on language, images, video, and action, with instruction understanding, world modeling, and action generation handled by one sequence model rather than separate subsystems [§4.3]. The model takes language + vision and drives 20-DoF dexterous hands within a 100 ms control loop [§Introducing Natura-1].
Pretraining corpus is 1,000,000 hours of egocentric human experience — 61,702,471 episodes, 317,842 tasks, 14,286 scenes — collected via a structured task-taxonomy pipeline backed by an agentic workflow that routes vendor-contributed data to underrepresented regions, plus large-scale internet ego videos; monocular, stereo, multi-camera, and wrist-mounted views [§4.1 Data].
In-house motion-labeling pipeline reports 41% MPJPE reduction vs existing open-source approaches on an internal benchmark with mocap ground truth [§4.1 “2. Motion Reconstruction”]. Hierarchical language annotation produced via an agentic workflow with shared per-episode context, keeping goal / intermediate-step / individual-action descriptions consistent [§4.1 “3. Language Annotation”].
Full-stack co-design is pitched as load-bearing: data infrastructure manages tens of PB in a unified lakehouse processing 350,000 hours/week; training runs stream batches directly from object storage with no shared file system and report MFU >50% + GPU utilization 98% on hundreds of GPUs with custom fused kernels [§4.1 “Scaling the Data System”, §4.2 Training]. Custom attention kernels deliver ~1,200 TFLOP/s average forward+backward — >6× tuned FlexAttention baseline — translating to >1.7× end-to-end training and 2× inference speedup in the evaluated configurations [§4.3 Kernel].
Inference stack is written from scratch, Rust-based for the orchestrator control plane and tensor runtime, with session state as a core responsibility (observations arrive asynchronously, each robot’s context persists across requests), separating control messages from bulk payloads, using GPU-resident state + reusable caches + captured execution graphs, and shared operator implementations across training and inference for numerical consistency — reports end-to-end inference latency within 100 ms [§4.2 Inference].
Benchmarking is explicitly head-to-head against π0.5 and GPT-6-Astra on a 12-dimension dexterous-manipulation benchmark. On Nature Will’s dual-arm ROKAE AR5 + Wuji hands platform, neither baseline scores: π0.5 fails even after fine-tuning on Nature Will’s dexterous-hand robot data, GPT-6-Astra fails to score within the 60-min per-task time limit. On ARX AC One parallel-gripper, after fine-tuning both baselines on gripper demos covering the same tasks as Natura-1’s dexterous-hand training set: π0.5 13.5, GPT-6-Astra 16.7 (with grippers); Natura-1 53.4 (Base), 59.1 (Agentic) with 20-DoF dexterous hands [§3.1].
Zero-shot instruction following holds under open-ended, spontaneous prompts recorded in day-to-day researcher interactions — including conditional and sequenced subtasks (“First use your left hand to pick up the toothbrush case and put it in the bin, then use your right hand to pick up the water bottle…”), with Natura-1 adjusting actions and continuing to pursue the instructed goal when objects move or execution is disrupted [§3.3].
Emergent recovery behaviors are reported in daily interactions that were not explicit evaluation targets: sliding a card over the edge of a yellow box to create a graspable overhang, using a rod to pull an object within reach, mid-air wrist rotation to steady a toppling stack — none specified in the prompt and none in the robot demonstrations [§3.4].
Scope limit on long-horizon demos: six of the demonstrations (marked †) use additional tuning on top of Natura-1; all other results share the same weights [§Real-World Capabilities footnote].

Method

Natura-1 is a 50B-parameter MoE transformer with a design kept close to standard — algorithmic innovations lie in the arrangement of multimodal tokens within a single sequence and in the causal structure of attention, keeping the backbone simple so it scales further [tweet 2109858487308255578]. Three stated optimization principles: how much the model learns per unit of compute, how much it learns per trajectory, and (continued in the truncated section).

The pretraining substrate is 1M hours of egocentric human experience processed into language-conditioned action trajectories by a pipeline that reconstructs dense 3D hand poses from monocular + stereo video, retargets human motion into robot-centric form, and attaches hierarchical language annotations. After filtering, this is the corpus the MoE is trained on. A limited post-training stage then adapts the pretrained world knowledge to a physical anthropomorphic dexterous hand — Nature Will’s framing is that anthropomorphic hands “offer the closest correspondence available today” to the human hand that the pretraining data is grounded in.

A key architectural claim in the blog’s §4.3 (model) is that actions are not separated out from world modeling — the same sequence model handles language, vision, and action, so instruction understanding, world prediction, and action generation share representations. The blog does not disclose the expert count, activation sparsity, action head structure, or vocabulary; evaluation is on a dual-arm platform (ROKAE AR5 with Wuji hands) with a 12-dimension dexterous-manipulation benchmark, plus zero-shot generalization and emergent-behavior studies.

Scaling evidence (§Pretraining data scales smoothly, tweet 2109858489724178928): with egocentric training data scaled from 44K to 1M hours, prediction MSE error on real-robot data falls by 67.7% without fine-tuning on robot data — Nature Will’s primary quantitative support for the “human-video scale is the lever” position.

Results

Benchmark head-to-head on 12-dimension dexterous-manipulation suite (§3.1): Natura-1 Base 53.4, Natura-1 Agentic 59.1 with 20-DoF dexterous hands on ROKAE AR5 + Wuji platform. π0.5 and GPT-6-Astra score zero on the same platform (π0.5 fine-tuned on Nature Will’s dexterous-hand data still fails; GPT-6-Astra does not complete tasks within 60 min). On ARX AC One parallel-gripper platform where both baselines were fine-tuned on matched-task gripper data: π0.5 13.5, GPT-6-Astra 16.7 — Natura-1’s dexterous-hand result (53.4) sits ~3.2–3.9× the fine-tuned baselines on parallel grippers.
Pretraining scaling law (§Pretraining data scales smoothly): 44K → 1M hours of egocentric human data lifts action prediction accuracy on real-robot data substantially; MSE error falls 67.7% without fine-tuning on robot data (tweet 2109858489724178928).
Infrastructure numbers: >50% MFU + 98% GPU utilization in production training; custom attention kernels ~1,200 TFLOP/s average (>6× FlexAttention); >1.7× end-to-end training speedup; 2× inference speedup; <100 ms end-to-end inference latency on physical hardware; 350K hours of ego data processed per week [§4.1, §4.2, §4.3].
Long-horizon tasks: six multi-minute single-take autonomous workflows (folding delicate deformable materials, calligraphy rendering with compliant brushstrokes, multi-stage tasks with mid-rollout tool switches). Six are explicitly flagged as using task-specific tuning on top of Natura-1; the remaining demonstrations share one checkpoint [§3.2, §Real-World Capabilities footnote].
Zero-shot instruction following: reproduces complex multi-arm sequencing from language (“First use your left hand… then your right hand…”), handles stacking (nest three cups → add orange → drape tissue over orange), and preserves the goal under disturbances [§3.3].
Emergent manipulation behaviors: card-overhang workaround for a graspable edge, rod-as-reach-tool for distant objects, wrist rotation on retry to steady toppling stacks — none in training demonstrations [§3.4].

Why it’s interesting

Natura-1 is the sharpest public datapoint yet on the 1M-hour egocentric-human-video pretraining position for VLA-style policies, directly comparable in scale to the only other filed instance on that frontier — Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models Dyna-2 — and significantly larger than EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data EgoScale (20,854 h) or Isaac 0.5: An Open-Weight Embodied Foundation Model Isaac 0.5 (375K h egocentric component). It joins the Human-to-Robot Retargeting “scale paid for by in-house motion-labeling pipeline” branch (41% MPJPE reduction vs open-source) rather than EgoScale’s 22-DoF joint channel or Translation-as-Bridging-Action’s 3-DoF wrist reduction, and keeps the full anthropomorphic 20-DoF channel paid for by scale.

Three positions Natura-1 stakes that are distinct from the existing VLA Models recipe-lever board: (1) unified 50B MoE for language + vision + world + action in one sequence model rather than VLM-backbone-with-action-head (π*0.6, Embodied-R1.5), WFM-plus-small-action-expert (µ₀, FLUX-mimic), or symmetric-I/O backbone (Rho-1); (2) full-stack co-design including custom Rust inference runtime at 100 ms end-to-end latency, pushing the “deployment-runtime axis” discussed in InstinctFlash — High-Performance Serving Runtime for Robotics Models InstinctFlash from aftermarket serving into co-designed training+inference infrastructure; (3) anthropomorphic-hand-as-closest-correspondence rather than parallel-gripper + UMI invariance (Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1) mimic, Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training XR-1) or embodiment-neutral mocap gloves (RewardAI OM-1 + Omnibody Hand launch — zero-shot cross-embodiment dexterous manipulation from human data (DexCap successor) RewardAI OM-1).

The π0.5 and GPT-6-Astra head-to-head is unusually sharp but also unusually confounded — the comparison is on dexterous hands where neither baseline was designed to operate, and the “baselines score 0” framing bakes in the dexterous-vs-gripper gap rather than isolating the data+architecture levers. Still, the parallel-gripper matched comparison (13.5 / 16.7 vs 53.4 / 59.1) gives a cleaner read, and the 1M-hour scaling MSE result (67.7% error reduction) is the first filed quantitative evidence that human-video pretraining transfers to zero-shot robot action prediction at this scale.