“Model training is not hard. But I need a lot of data… to train my robots.”
We hear this from robotics labs constantly.
Software AI learned from the internet. Robots must learn from the real world, where usable data is still scarce.
That’s why THOR starts with data.
We hear this from robotics labs constantly.
Software AI learned from the internet. Robots must learn from the real world, where usable data is still scarce.
That’s why THOR starts with data.
1
1
13
October was THOR’s first month live on Bittensor SN112.
• First robotics labelling contracts signed
• 3 annotation firms onboarded
• Invite-only orderbook
• talorobotics.ai launched
November is about delivery: signed contracts → accepted work → subnet revenue.
• First robotics labelling contracts signed
• 3 annotation firms onboarded
• Invite-only orderbook
• talorobotics.ai launched
November is about delivery: signed contracts → accepted work → subnet revenue.
2
3
22
THOR Signals · Capital
@Figure_robot tried to buy robot training data. Vendors couldn’t hit the throughput, diversity or quality it needed.
So it built its own pipeline, and in August committed $1B+ to data and compute over 12 months.
Meanwhile, data suppliers are raising:
→ led by Sequoia, with Nvidia and M12 joining.
→ @MaxinsightsAI: $30M seed led by Peak XV. 2M+ hours of human activity video captured, with 1.5M rated “model-ready” under its internal scoring system.
As more capital flows into collection, more data needs verifying, and that’s where things slip.
Pantheon audited 3,546 episodes across nine open robotics datasets and found 27% had at least one medium or high severity issue. That's open data, not these suppliers’, but it shows how easily problems get through.
Figure’s answer: keep filtering, fraud review and deduplication in-house.
Our view: an open market for robotics data needs verification buyers trust enough to cut back their own checking.
THOR’s pilot is testing whether independent verification can match buyers’ own quality calls, and cut the time they spend checking.
@Figure_robot tried to buy robot training data. Vendors couldn’t hit the throughput, diversity or quality it needed.
So it built its own pipeline, and in August committed $1B+ to data and compute over 12 months.
Meanwhile, data suppliers are raising:
→ led by Sequoia, with Nvidia and M12 joining.
→ @MaxinsightsAI: $30M seed led by Peak XV. 2M+ hours of human activity video captured, with 1.5M rated “model-ready” under its internal scoring system.
As more capital flows into collection, more data needs verifying, and that’s where things slip.
Pantheon audited 3,546 episodes across nine open robotics datasets and found 27% had at least one medium or high severity issue. That's open data, not these suppliers’, but it shows how easily problems get through.
Figure’s answer: keep filtering, fraud review and deduplication in-house.
Our view: an open market for robotics data needs verification buyers trust enough to cut back their own checking.
THOR’s pilot is testing whether independent verification can match buyers’ own quality calls, and cut the time they spend checking.
1
4
20
If everyone has access to better AI models, what do contributors compete on?
How well they use them.
Models like GPT-6 Astra and NVIDIA’s Flora are getting much better at labelling robotics video. NVIDIA estimates inference compute costs below $1 per hour of footage for Flora.
a has to meet timing, description and quality requirements. Verification and rework carry costs too.
On Bittensor Subnet 112, THOR is designed to tie rewards to verified output, giving contributors an incentive to improve their models, tools and workflows. A better workflow raises the bar for everyone.
The test for THOR is whether that competition delivers reliable robotics training data at a lower total cost.
Better AI expands what’s possible. Competition makes it count.
How well they use them.
Models like GPT-6 Astra and NVIDIA’s Flora are getting much better at labelling robotics video. NVIDIA estimates inference compute costs below $1 per hour of footage for Flora.
a has to meet timing, description and quality requirements. Verification and rework carry costs too.
On Bittensor Subnet 112, THOR is designed to tie rewards to verified output, giving contributors an incentive to improve their models, tools and workflows. A better workflow raises the bar for everyone.
The test for THOR is whether that competition delivers reliable robotics training data at a lower total cost.
Better AI expands what’s possible. Competition makes it count.
0
7
25
500,000+ hours of raw video behind one robotics model. A city subsidising companies' data purchases.
China's robotics data work is turning into a supply chain.
This is THOR Signals, our new series on the robotics data market, starting with China.
For context, China accounted trial robot installations in 2025. Within that industry, three roles are taking shape:
• @beingbeyond_: model development
BeingBeyond builds models that help robots understand their surroundings and perform physical tasks. For Being-H0.8, the model behind those 500,000+ hours, raw video was filtered, reconstructed and reviewed before training.
• @PaXiniAI: specialist data supply
PaXini makes tactile sensors and training data combining vision, touch and movement. BeingBeyond names it an important data partner for Being-H0.8.
• @LightwheelAI: connecting training with deployment
Lightwheel supplies simulation, demonstration data and evaluation, and reported about $100M in Q1 2026 orders. At automaker Geely, factory demonstrations feed training, and testing shows where more data is needed.
Government is shaping demand too. Hangzhou's 2026–2027 action plan subsidises companies' purchases of data products and services (up to 30%, capped at RMB300,000). The national pilot base there invited outside data partners on 30 September across 25 task scenarios.
What we don't know yet: how much teams outsource, and which services they buy repeatedly.
For THOR, that's the question: which tasks do teams commission externally, and why choose a specialist over building in-house?
More from THOR in China to come.
THOR Signals · China
China's robotics data work is turning into a supply chain.
This is THOR Signals, our new series on the robotics data market, starting with China.
For context, China accounted trial robot installations in 2025. Within that industry, three roles are taking shape:
• @beingbeyond_: model development
BeingBeyond builds models that help robots understand their surroundings and perform physical tasks. For Being-H0.8, the model behind those 500,000+ hours, raw video was filtered, reconstructed and reviewed before training.
• @PaXiniAI: specialist data supply
PaXini makes tactile sensors and training data combining vision, touch and movement. BeingBeyond names it an important data partner for Being-H0.8.
• @LightwheelAI: connecting training with deployment
Lightwheel supplies simulation, demonstration data and evaluation, and reported about $100M in Q1 2026 orders. At automaker Geely, factory demonstrations feed training, and testing shows where more data is needed.
Government is shaping demand too. Hangzhou's 2026–2027 action plan subsidises companies' purchases of data products and services (up to 30%, capped at RMB300,000). The national pilot base there invited outside data partners on 30 September across 25 task scenarios.
What we don't know yet: how much teams outsource, and which services they buy repeatedly.
For THOR, that's the question: which tasks do teams commission externally, and why choose a specialist over building in-house?
More from THOR in China to come.
THOR Signals · China
3
9
28
Robotics training data needs to reflect the environments where robots will work.
@haitzu on THOR’s open data marketplace at Exploit Summit:
“If these robots need to be everywhere, what we need to do is to get as many people from around the world contributing that data.”
@haitzu on THOR’s open data marketplace at Exploit Summit:
“If these robots need to be everywhere, what we need to do is to get as many people from around the world contributing that data.”
0
6
22
The embodied AI data market is forecast to grow from an estimated $536M in 2025 to $4.9B by 2032. Roughly 9x in seven years.
But volume alone isn't enough. A demonstration recorded for one robot arm can be useless to a lab training a different one. Labs need data cut into elled with language instructions and checked for consistency, all to their own spec.
THOR starts with what the lab needs, then turns those requirements into paid annotation and verification tasks for providers on Bittensor. Outputs are checked against the lab’s specifications.
We're building so labs can scale their data operations without coordinating every provider themselves.
But volume alone isn't enough. A demonstration recorded for one robot arm can be useless to a lab training a different one. Labs need data cut into elled with language instructions and checked for consistency, all to their own spec.
THOR starts with what the lab needs, then turns those requirements into paid annotation and verification tasks for providers on Bittensor. Outputs are checked against the lab’s specifications.
We're building so labs can scale their data operations without coordinating every provider themselves.
1
9
22
Robots have no equivalent of the internet to learn physical work from.
THOR’s mission is to build that foundation in the open.
Annotation is where that journey begins.
Read our manifesto:
thormachines.ai/manifesto/
THOR’s mission is to build that foundation in the open.
Annotation is where that journey begins.
Read our manifesto:
thormachines.ai/manifesto/
4
12
31
