First general foundation model for the physical world
5796535827K209
Tags
B2D (developers)Product launch500K-1MB2BDemoUS
TLVC Rating
2.80
Hook
Editing / Creativity
Copy
Sentiment of launch
Distribution strategy
Community Rating
No ratings yet
to rate this launch.
Natura-1 is a foundation model from Nature Will that takes natural language instructions and vision input and drives dual-arm robots with dexterous hands. It is a 50-billion-parameter mixture-of-experts model pretrained on more than one million hours of egocentric human experience, and it drives 20-DoF dexterous hands within a 100-millisecond control loop. The audience is robotics teams and physical AI researchers who have been limited to either general policies for simple grippers or task-specific systems for dexterous hands, and this launch is Nature Will's first public research preview of a single model that attempts both at once.
Natura-1 follows complex natural-language instructions zero-shot, reasoning through spatial constraints, multi-object relationships, and negative conditions , and the demos cover long-horizon workflows such as folding deformable fabric, rendering calligraphy with compliant brushstrokes, and switching tools mid-task. The thesis behind the data is explicit. Robot data collection is bounded by robot capability, so Natura-1 instead learns from human experience, with the curated corpus comprising one million hours of language-conditioned action trajectories recovered as dense 3D hand poses from monocular and stereo video.
Nature Will is a physical AI lab dedicated to building physical AGI , headquartered in San Francisco and led by founder and CEO Ning Ding. The team built the training and serving stack in-house, including a distributed engine co-designed with the MoE architecture and a Rust inference orchestrator that holds the 100-millisecond observation-to-action loop on physical hardware, which matters because it signals the lab intends to keep scaling data, model size, and robot deployments from the same foundation rather than retraining per task.
Comments (15)
Sign in to join the discussion.
P
Priya Vaswani20h ago
One million hours of what exactly? If it's 900k hours of people folding laundry we're going to have the world's most overqualified sock sorter.
T
Tomas Lindqvist20h ago
The pacing on that launch video is suspiciously good. Who did the edit, because the cut from gripper to voiceover is doing more work than the model probably is.
D
Dilnoza Karimova19h ago
The thumbnail is doing that beige-on-beige thing where the hand blends into the background. Tighten the contrast and that frame becomes iconic.
Z
zane20h ago
Every demo video is shot in a bright warehouse with one cube and one banana. Call me when Natura-1 survives my kitchen counter at 11pm.
N
Noemi Astra19h ago
The cube pick-and-place has been the robotics equivalent of 'hello world' for a decade. Show me Natura-1 unloading a dishwasher and I'll believe the 'general' part.
D
Devika R.20h ago
We prototyped something adjacent at Meta Reality Labs in 2019 and killed it because the eval harness was a nightmare. Curious how you're measuring 'zero-shot' here without it becoming vibes-based.
F
Fatimah Zahra20h ago
Does Natura-1 ship with SSO and SOC2 Type II, or are we doing the usual 'enterprise tier coming soon' dance? Procurement is already asking me and the model isn't even in a box yet.
S
Sven B.19h ago
'First general foundation model for the physical world' is a bold claim given how many labs are quietly training the exact same thing in a Palo Alto garage.
M
mireille okafor19h ago
Tagline rewrite, no charge: 'Natura-1. Words in, work out.' You can venmo me later.
H
Hiroshi Vale19h ago
This is genuinely the last robotics foundation model that will be shipped without an agent layer on top. Within a year nobody will prompt these things directly, they'll just negotiate with them.
K
Kwame O.19h ago
If this really is the last pre-agent product, you're going to look prescient or very embarrassed in 18 months. Either way the thread will age interestingly.
J
Johanna Keller19h ago
Been watching this team cook for months, so proud of the reveal. The way MoE routing was explained in the thread will be in a lecture slide by Tuesday.
O
Oluwaseun A.19h ago
Where does the one million hours of pretraining footage live, and under which jurisdiction? If any of it touches EU subjects we need to talk about Article 6 before you touch a European warehouse.
K
kostya19h ago
Serious question: how many people on the team? Because if a sub-20 crew pretrained a 50B MoE I'm quitting my job tonight and buying a GPU on marketplace.
R
Rohit Mehrotra19h ago
Interesting that there's no mention of a decentralized data layer for the hours of robot teleop. Feels like a missed coordination mechanism when operators could be rewarded for contributing.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
nature will
First general foundation model for the physical world
First general foundation model for the physical world
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Natura-1 is a foundation model from Nature Will that takes natural language instructions and vision input and drives dual-arm robots with dexterous hands. It is a 50-billion-parameter mixture-of-experts model pretrained on more than one million hours of egocentric human experience, and it drives 20-DoF dexterous hands within a 100-millisecond control loop. The audience is robotics teams and physical AI researchers who have been limited to either general policies for simple grippers or task-specific systems for dexterous hands, and this launch is Nature Will's first public research preview of a single model that attempts both at once. Natura-1 follows complex natural-language instructions zero-shot, reasoning through spatial constraints, multi-object relationships, and negative conditions , and the demos cover long-horizon workflows such as folding deformable fabric, rendering calligraphy with compliant brushstrokes, and switching tools mid-task. The thesis behind the data is explicit. Robot data collection is bounded by robot capability, so Natura-1 instead learns from human experience, with the curated corpus comprising one million hours of language-conditioned action trajectories recovered as dense 3D hand poses from monocular and stereo video. Nature Will is a physical AI lab dedicated to building physical AGI , headquartered in San Francisco and led by founder and CEO Ning Ding. The team built the training and serving stack in-house, including a distributed engine co-designed with the MoE architecture and a Rust inference orchestrator that holds the 100-millisecond observation-to-action loop on physical hardware, which matters because it signals the lab intends to keep scaling data, model size, and robot deployments from the same foundation rather than retraining per task.
Comments (15)
One million hours of what exactly? If it's 900k hours of people folding laundry we're going to have the world's most overqualified sock sorter.
The pacing on that launch video is suspiciously good. Who did the edit, because the cut from gripper to voiceover is doing more work than the model probably is.
The thumbnail is doing that beige-on-beige thing where the hand blends into the background. Tighten the contrast and that frame becomes iconic.
Every demo video is shot in a bright warehouse with one cube and one banana. Call me when Natura-1 survives my kitchen counter at 11pm.
The cube pick-and-place has been the robotics equivalent of 'hello world' for a decade. Show me Natura-1 unloading a dishwasher and I'll believe the 'general' part.
We prototyped something adjacent at Meta Reality Labs in 2019 and killed it because the eval harness was a nightmare. Curious how you're measuring 'zero-shot' here without it becoming vibes-based.
Does Natura-1 ship with SSO and SOC2 Type II, or are we doing the usual 'enterprise tier coming soon' dance? Procurement is already asking me and the model isn't even in a box yet.
'First general foundation model for the physical world' is a bold claim given how many labs are quietly training the exact same thing in a Palo Alto garage.
Tagline rewrite, no charge: 'Natura-1. Words in, work out.' You can venmo me later.
This is genuinely the last robotics foundation model that will be shipped without an agent layer on top. Within a year nobody will prompt these things directly, they'll just negotiate with them.
If this really is the last pre-agent product, you're going to look prescient or very embarrassed in 18 months. Either way the thread will age interestingly.
Been watching this team cook for months, so proud of the reveal. The way MoE routing was explained in the thread will be in a lecture slide by Tuesday.
Where does the one million hours of pretraining footage live, and under which jurisdiction? If any of it touches EU subjects we need to talk about Article 6 before you touch a European warehouse.
Serious question: how many people on the team? Because if a sub-20 crew pretrained a 50B MoE I'm quitting my job tonight and buying a GPU on marketplace.
Interesting that there's no mention of a decentralized data layer for the hours of robot teleop. Feels like a missed coordination mechanism when operators could be rewarded for contributing.