Skip to content

THOR Group / 01–08

Build physical intelligence in the open.

The world will teach machines to act. The people who do that work should share in the value they create.

Written by
Jack Ai-Leung
Published

Scroll to read

The moment

Intelligence is entering the physical world.

Long before robots entered industry, they lived in our imagination. R2-D2. WALL-E. The Iron Giant. Machines that could move beside us, extend our abilities and help civilisation accomplish more than people could achieve alone. That vision is coming within reach as AI moves beyond screens and learns to act in the physical world.

These machines could change how we work and live. They could take on dangerous jobs, support ageing populations and strengthen manufacturing. The consequences reach far beyond another software cycle.

Realising that potential will take years of engineering, research and learning in the physical world. The choices we make now will determine who gets to build this future, who benefits from it and who holds the power.

Physical AI should be built with the world, not merely delivered to it. If the experience that teaches machines to act stays inside a handful of private laboratories, those organisations will set the terms for everyone else. Researchers, builders and communities everywhere should be able to contribute and develop capabilities of their own.

THOR Group brings together an open network of contributors, verification and incentives through SN112, THOR’s subnet on Bittensor, and relationships with Chinese manufacturers and robotics researchers around the world. We are building businesses that connect these capabilities: producing training data, supporting open models and putting robots to work.

We begin with annotation: turning existing video into structured training data by identifying actions and marking when they begin and end. Our first planned release will make that dataset openly available for researchers to inspect, evaluate and build on.

Section 01: The physical world was never archived.

For machines to work alongside us, they must learn from physical experience. Much of that experience has never been recorded in a form they can learn from.

  • Language models inherited a record that already existed: books, code, images and conversations had accumulated for decades before anyone trained on them. Robots have no equivalent archive of physical action. There is no internet-scale record of how a hand peels a carrot, threads a cable or steadies a tote bag as it deforms.
  • The knowledge a robot needs lives in human hands, factory routines, household improvisation and machine failures. Teaching it to grasp, pour, fold or recover from a slip requires someone to capture that experience deliberately. Simulation can expand what we learn from it, but it must remain grounded in how the physical world behaves.
  • We believe the shortage of suitable training data is the central constraint on progress: shared datasets1 and policies trained across diverse tasks2 show the value of learning from a wider range of experience. Model architectures will change. The need for diverse, well-prepared demonstrations will remain.
  • No single institution can capture all the tasks, bodies and environments machines will encounter. That takes coordination: people in different places contributing what they know, with their contributions recognised and rewarded.

The record robots need will be built task by task, by people across the world.

Supporting research (4)
  1. 01

    Open X-Embodiment

    This is the largest open real-robot dataset there is. Pools 60 datasets from 34 labs into 1M+ trajectories across 22 embodiments, and was explicitly framed as testing whether a “consolidation” of pretrained models like the one in NLP and vision can happen in robotics.

  2. 02

    Physical Intelligence π0.5

    Demonstrated open-world generalisation, cleaning entirely new kitchens and bedrooms it had never seen, using co-training on heterogeneous data: multiple robots, web data, and high-level semantic prediction.

  3. 09

    Ego4D

    About 3,670 hours of egocentric daily-life video from 931 camera-wearers across 74 locations in 9 countries.

  4. 10

    Ego-Exo4D (V2)

    1,286 hours of time-synced egocentric and exocentric video of skilled activities from more than 800 participants across 131 scene contexts, with 7-channel audio, IMU, eye gaze and SLAM cameras.

Section 02: Progress should become a shared foundation.

Capturing physical experience is the first step. The next is making that knowledge available to people who can use it, improve it and take robotics further.

  • Openness determines who gets to build: a university should be able to adapt a model to an affordable robot. A startup should be able to develop a manufacturing application. A local team should be able to work on agricultural conditions its own community understands. Access to data, usable weights and reproducible hardware gives those teams a starting point.
  • Open models16, community datasets39 and published hardware designs37 already show what this approach makes possible. Teams can examine one another’s work, adapt it and contribute improvements. Progress can accumulate across organisations.
  • Physical experience can train both robot policies and world models. Robot policies19 learn to turn observations and instructions into actions. World models42 learn how environments change and can predict what a robot might see after it acts. These approaches offer ways to learn from demonstrations and test possible outcomes before a robot moves.
  • A closed policy keeps capability within one company. An open policy lets other teams build on it. Teams can run an open model locally, adapt it under its licence and continue using downloaded weights if the original provider exits. Training code, documentation and clear reuse rights help them put it to work.

Fig 01

Robotics should have many independent centres of development.Provider-controlled modelOpen-weight modelOne providercontrols the modelOpen weightsanyone can run themTeamTeamTeamTeamA university adapts it to an affordable robot armA startup trains it for a manufacturing taskA research lab evaluates failures and publishes themA local team adapts it to agricultural conditionsDependent on one providerImprovements flow both waysThe provider-controlled shape — One provider trains the model and decides which services it offers. Every team downstream reaches it on those terms, and their work continues only while that access does.The open-weight shape — A university adapts the model to an affordable robot arm, a startup trains it for a manufacturing task, a research lab evaluates failures and publishes improvements, and a local team adapts it to agricultural conditions. Each is a centre of development rather than a customer, and what each learns can flow back.Provider-controlledOpen-weightAccessThrough the provider’s offered services and termsDownload and run under the model’s licenceAccess — A provider decides which services it offers and on what terms. Open weights are downloaded and run under the licence the model carries, with no permission step in between.AdaptationWithin the tools and permissions offeredFine-tune for local robots, tasks and environmentsAdaptation — Adapting a hosted model means working inside the tools and permissions the provider offers. Open weights can be fine-tuned for the robots a team owns, on data that team collected itself.ContinuityDepends on continued provider accessDownloaded weights can remain availableif the original provider exitsContinuity — A hosted model stops when the service does. A downloaded set of weights keeps working whatever happens to the organisation that trained it, which is what makes a local capability durable.DevelopmentCore model changes controlled by its ownerMany teams develop and share derivativesDevelopment — When the core model belongs to one owner, that owner decides what it becomes. Open weights let several teams develop derivatives at once and evaluate each other’s work.Shared weights allow teams around the world to develop local capabilities from a common foundation, reducingdependence on any single model provider.

Click any part for the detail

Sourcesπ0.52LeRobot38

  • An independent data network should give university researchers free access to public datasets and research resources. Commercial customers may also commission private work; open participation can accommodate both.

Each team’s progress should give the next team a stronger starting point.

Supporting research (7)
  1. 16

    Physical Intelligence π0

    Released as open weights and code: a flow-matching VLA built on PaliGemma, plus π0-FAST, a tokenizer that trains generalist policies 5x faster.

  2. 17

    NVIDIA GR00T N1

    Released as “the world’s first open, fully customizable foundation model” for humanoids, with a dual-system architecture: a System 2 VLM for reasoning and a System 1 diffusion transformer generating actions at up to 120 Hz.

  3. 19

    OpenVLA

    An open-source vision-language-action model with released checkpoints and code. Its paper reports pre-training on diverse real-robot demonstrations and fine-tuning for new settings, illustrating how broad pre-training can be followed by task-specific adaptation.

  4. 20

    Hugging Face SmolVLA

    A lightweight 0.5B open VLA trained entirely on LeRobot community datasets rather than proprietary data, deliberately using messy, realistic, crowd-sourced interactions.

  5. 42

    World models for physical AI: NVIDIA Cosmos

    NVIDIA’s Cosmos research demonstrates world models that use observations and controls to predict future video, with applications in planning, evaluation and data generation. These models illustrate how recorded physical experience can support learned representations of how environments change; they do not establish that world models replace physics engines or that any one video perspective is always best.

  6. 39

    LeRobot community datasets

    SmolVLA was trained on 487 curated community datasets under the lerobot tag. Community Dataset v1 alone was 128 datasets from 55 contributors: 11.1K episodes, 5.1M frames, 46.9 hours. Later community sets drew from 235 contributors across 46+ embodiments, proving decentralised open data collection works.

  7. 37

    Open robot hardware, published designs

    Pollen Robotics grew out of Inria work on Poppy and is now part of Hugging Face; its Amazing Hand publishes a bill of materials, printable designs, assembly instructions and demo programs for a parts cost below €200. The ORCA Hand, from ETH Zurich’s Soft Robotics Lab, is a 17-degree-of-freedom platform with a reported $2,000 bill of materials, though its v2 touch-sensing module stays proprietary. Enactic’s OpenArm publishes CAD, firmware, control code and simulation tools with the hardware repository under CERN-OHL-S-2.0. The Open Dynamic Robot Initiative, a Max Planck, LAAS-CNRS and NYU collaboration, publishes drawings, electronics and software for modular actuators and legged platforms under BSD 3-Clause. Berkeley Humanoid Lite publishes schematics, bill of materials, firmware, control software and training code for a 3D-printed humanoid costing about $4,300 in the United States.

Section 03: More footage is not more useful training data.

  • A lab buys an answer to a particular problem: consider a team teaching a robot to sort parts. It needs examples matched to its camera and gripper, with labels that distinguish a successful grasp from a failed one. A large library of household video may not answer that brief. Finding a supplier, explaining the task and checking the result can become a project of its own.
  • Pre-training and post-training serve different purposes. Broad demonstrations help establish a model’s capabilities during pre-training. Post-training can then adapt those capabilities to a particular robot or task, using targeted demonstrations19. That distinction matters for THOR: the value of a dataset depends on what the researcher needs it to improve.
  • A human demonstration has to be adapted to what a robot can see and do. A camera on a person’s head gives a different view from one on a robot’s wrist, while fingers and grippers handle objects differently. Video can also leave out information the robot needs. Footage of someone inserting a plug may hide the contact point and contain no measurement of the force applied.
  • The collection method should follow the training need. Teleoperation captures actions on the target robot, while UMI-style grippers5 offer a more portable way to record demonstrations. Human demonstrations provide another source of experience, captured through gloves, tracking systems or video recorded from a person’s perspective. Making that experience usable for robot learning requires careful tracking and preparation, as EgoMimic7 and EgoDex8 illustrate. The right method balances data quality, compatibility with the target robot and the cost of collection.
  • Simulation can extend that process, using tools such as the Newton physics engine32. A captured scene can be reconstructed and varied to generate more training episodes. AgiBot’s 2026 challenge36 combines real-robot datasets with digital-twin simulation for model evaluation. The test is whether performance in simulation carries over to the physical world: the sim-to-real gap. Cloth manipulation remains challenging to simulate faithfully13, although specific tasks have shown promising transfer to real robots.

Fig 02

Different ways of collecting robot data balance scale, realism and compatibility with the robot.Easier to collectFurther from robot actionsHarder to collectCloser to robot actionsCovered bySimulationScalable robot-action data; transfer depends on simulation fidelity and training design.THOR SimulationSimulation — Simulation scales data generation but approximates physics. Cloth remains challenging to simulate faithfully, although simulation-trained methods have demonstrated useful results on specific tasks such as cloth smoothing. Transfer still needs real-world evaluation. THOR Simulation rebuilds a captured scene and varies it, so one capture becomes many episodes.Egocentric human videoDiverse real-world behaviour; robot actions must be inferred or learned indirectly.THOR DataEgocentric human video — The datasets already exist at scale, and the behaviour is natural with no embodiment constraints. Hand-pose tracking from egocentric video is less solved than the field assumes, and precision collapses in contact-rich tasks such as plug and unplug. THOR Data is where this gets ordered, annotated and verified.UMI-style handheld gripperA promising option for scalable parallel-jaw manipulation data.THOR HardwareUMI-style handheld gripper — Standardised, 3D-printable and postable to any collector, and the paradigm now extends to dexterous hands and tactile sensing. THOR Hardware builds the capture kits on the stack every lab already runs.Gloves and exoskeletonsCapture dexterous motion; require mapping to the target robot hand.THOR HardwareGloves and exoskeletons — Gloves capture real hands but interfere with natural motion and add cognitive load, changing the behaviour you are trying to capture. Mapping that motion to a target hand is a hardware problem, which is why THOR Hardware owns it.TeleoperationDirect demonstrations on robot hardware; requires robot access and operator time.THOR DataTeleoperation — Precise joint-to-joint mapping and the current gold standard for post-training, but it overfits to specific tasks and is expensive in human hours. THOR Data buys those hours on the network rather than on a payroll.Autonomous rollouts and interventionsDeployment experience for policy improvement.THOR DeploymentAutonomous rollouts and interventions — Once a policy is fielded, its own rollouts and the moments a human has to step in are training data. It needs robots already deployed, which is why THOR Deployment is the unit that reaches it, and why it compounds fastest.

Click any part for the detail

SourcesUMI5DexUMI6

  • For THOR, collection starts with the researcher’s brief: the task, target hardware, failures to address and criteria for accepting the data. Annotation follows those requirements. Results on the robot then guide the next round of collection.
Supporting research (10)
  1. 05

    UMI (Universal Manipulation Interface)

    A handheld 3D-printed gripper with a single wrist-mounted GoPro as its only sensor. Explicitly designed to be portable and reproducible so “researchers and enthusiasts should be able to consistently build UMI grippers,” and it enables zero-shot transfer across robot platforms.

  2. 06

    DexUMI, FastUMI, ViTaMIn, FreeTacMan

    Extend the handheld-interface paradigm to dexterous hands and add visuo-tactile sensing. FastUMI logs 6-DoF end-effector poses directly via RealSense T265 to avoid heavy offline SLAM.

  3. 07

    EgoMimic

    A Georgia Tech framework, published at ICRA 2025, that scales manipulation with egocentric human video paired with 3D hand tracking, and learns one policy over human and robot demonstrations together rather than extracting intent from the video alone. It reports that an added hour of hand data is worth more than an added hour of robot data.

  4. 08

    EgoDex

    829 hours of 1080p egocentric video from Apple Vision Pro, 338,000 episodes across 194 tabletop tasks, with 3D hand and finger pose tracked at capture time by on-device SLAM and calibrated cameras, plus camera extrinsics and language annotation. The tracking is what makes the video usable, and it was recorded, not recovered afterwards.

  5. 11

    EPIC-KITCHENS-100

    100 hours of densely annotated egocentric kitchen video. The VISOR extension adds 272,000 semantic masks and 9.9M dense masks.

  6. 12

    DROID

    An in-the-wild robot manipulation dataset with 76,000 demonstration trajectories, 350 hours of interaction data, 564 scenes and 86 tasks. Its 13 participating institutions used the same robot hardware setup to streamline collection and improve portability, and the project publishes a guide for reproducing that setup.

  7. 19

    OpenVLA

    An open-source vision-language-action model with released checkpoints and code. Its paper reports pre-training on diverse real-robot demonstrations and fine-tuning for new settings, illustrating how broad pre-training can be followed by task-specific adaptation.

  8. 32

    Newton physics engine

    NVIDIA, Google DeepMind and Disney Research are building Newton, an open-source GPU-accelerated physics engine built on Warp and integrating MuJoCo Warp, managed by the Linux Foundation, achieving over 70x simulation speedups.

  9. 36

    AgiBot real data and simulation workflows

    AgiBot’s ICRA 2026 challenge provides real-robot datasets and uses Genie Sim 3.0 digital-twin environments for model evaluation, with simulation and real-robot stages. Its world-model track trains video generation models using robot observations and action signals. This is a concrete example of combining real data and simulation, rather than proof that the whole field follows one pipeline.

  10. 13

    Cloth simulation and real-world transfer

    Benchmarking the Sim-to-Real Gap in Cloth Manipulation compares four simulators against real cloth manipulation tasks, examining fidelity, computational time and stability. Learning Visible Connectivity Dynamics for Cloth Smoothing demonstrates zero-shot transfer from simulation to a real robot on a specific cloth-smoothing task. These results support a task-specific assessment of simulation, not a blanket claim that it is uncompetitive or that a fixed amount of teleoperation is always preferable.

Fig 03

Data collection driven by research demandStarting PointCollectionPreparationTrainingNext IterationBefore: Buying available dataWith THOR: Commissioning required dataSupplier collectsfootage to sellLab buys an existingdatasetLab cleans, labelsand adapts the dataLab discovers gapsduring trainingLab searches foranother suitabledatasetLab defines atraining needTHOR recruits andrewards workersagainst the specContributions arechecked against theagreed requirementsLab trains andevaluates on itstarget robotResults guide thenext collection roundFailures guide collection · incentives sustain supplyStarting Point — A supplier decides what to film and then looks for a buyer, so the footage exists before anyone has said what it is for. Commissioned collection starts from the lab: the task, the hardware and the failure it wants fixed are named first.Starting Point — A supplier decides what to film and then looks for a buyer, so the footage exists before anyone has said what it is for. Commissioned collection starts from the lab: the task, the hardware and the failure it wants fixed are named first.Collection — Buying an existing dataset means taking whatever was captured, at whatever calibration and framing the supplier chose. Against a specification, THOR recruits the people who can meet it and pays them for output that does.Collection — Buying an existing dataset means taking whatever was captured, at whatever calibration and framing the supplier chose. Against a specification, THOR recruits the people who can meet it and pays them for output that does.Preparation — Buying data leaves the lab to clean, label and adapt it before it can test whether the data helps. Checking contributions against the agreed requirements moves that work off the lab.Preparation — Buying data leaves the lab to clean, label and adapt it before it can test whether the data helps. Checking contributions against the agreed requirements moves that work off the lab.Training — Gaps found during training are expensive, because the data that would close them was never collected. Evaluating on the target robot turns the gap into a specification for the next round rather than a dead end.Training — Gaps found during training are expensive, because the data that would close them was never collected. Evaluating on the target robot turns the gap into a specification for the next round rather than a dead end.Next Iteration — The search for another suitable dataset starts the whole problem again. With THOR, what training shows becomes the brief for the next collection round.Next Iteration — The search for another suitable dataset starts the whole problem again. With THOR, what training shows becomes the brief for the next collection round.

Click any part for the detail

SourcesEgoMimic7EgoDex8UMI5

Section 04: Verification makes the market possible.

  • The buyer sets the requirements: source footage, labels, task boundaries and a quality bar. THOR’s proposed market turns that brief into work that contributors compete to fulfil. Accepted output returns to the customer, and performance determines rewards.
  • Contributors may be annotation firms, specialised model teams or people using AI tools with expert review. They can participate from wherever they have the skills to fulfil the brief. Paying for output that meets the agreed standard gives each a reason to improve its tools and methods.
  • SN112 supplies the incentive layer: Bittensor gives THOR a way to reward production and verification while recruiting contributors. Validators check submissions against the customer’s criteria. They use models and hidden control samples with known answers to measure annotation quality. Those controls must reflect the real tasks and remain difficult to identify.

Fig 04

Miners produce annotations. Hidden quality checks help determine their rewards.Customerfootage and rubricValidatordistribute work,inject goldMinersannotate the footageValidatorverify against gold,judge accuracyEmissionspaid on accepted workGold: known-answer control sampleshidden among tasks to check annotation qualityThe order — A customer submits raw footage with an acceptance rubric and an hourly rate.Distribution — The validator distributes the work across miners and injects known-answer control samples into it.Annotation — Miners annotate the footage. Control samples are mixed into the workload without being identified to them; keeping those samples difficult to detect is part of the verification challenge.Verification — Validators run a blend of models over what comes back, reconcile it against the known answers and judge accuracy before anything is paid.Payment — Emissions flow to the miners whose work was accepted. The aim is to automate verification and payment; how much human review is needed to meet the quality bar remains to be tested.Gold, the control samples — Control samples with known answers are mixed into the workload without identifying them to miners. They provide a check on annotation quality, but their usefulness depends on remaining hard to identify and reflecting the tasks being scored. They do not by themselves establish that the system cannot be gamed.

Click any part for the detail

  • Reliable annotations at lower cost are the hypothesis. The first release must test quality and total cost, including verification and human review. Its performance will matter more than the elegance of the mechanism.
  • The first release will use existing egocentric footage, such as EPIC-KITCHENS11, to develop and evaluate the process. The annotated dataset will be released openly. A generalisable annotation model is a later ambition.

Trust and rewards must be earned through the quality of the work delivered.

Supporting research (1)
  1. 11

    EPIC-KITCHENS-100

    100 hours of densely annotated egocentric kitchen video. The VISOR extension adds 272,000 semantic masks and 9.9M dense masks.

Section 05: THOR Group will be Hugging Face for robotics.

An open ecosystem lasts when people can build on its work and earn from contributing to it.

  • Hugging Face brought models, datasets, tools and a global community into one open ecosystem. LeRobot38 already extends that approach to robotics. THOR’s ambition is to bring that scale of participation to producing physical interaction data, rewarding the people who collect, annotate and verify it.
  • We are building a market where that work is straightforward to commission and dependable to receive. Over time, a lab should be able to come to THOR for the data, open models and capture hardware a project needs, with the group coordinating suppliers and delivery. That gives customers one place to turn as their requirements grow.
  • Openness can support a commercial business. Customers pay for accepted work, and THOR can earn from defining requirements, matching orders to contributors, verifying quality and coordinating delivery. Published datasets can remain freely accessible while customers pay for work tailored to their needs.
  • Token incentives can help establish supply before commercial demand reaches scale. Customer revenue can then support continued production and the network’s open research resources. The business depends on delivering data that customers return for, at a price that supports the work.
  • Robotics data vendors15 already sell services in this market. Their reported funding shows investor interest in expanding the supply of training data and the tools needed to produce it.

Fig 05

Top data vendors to robotics have raised >$500mData vendorReported fundingLightwheelChina$280m+Lightwheel, $280m+ — Simulation assets, synthetic training data, egocentric human data and evaluation infrastructure. Two disclosed financings of about $140m each landed in March and June, and it combines simulated environments with human demonstrations through its EgoSuite product.XDOFUnited States$70mXDOF, $70m — Production-scale robotics datasets, collection systems and the tools around them, disclosed at launch in June with Thrive, Spark, a16z, Lux and WndrCo behind it. Its focus is robot teleoperation, with human-motion collection in development, which puts it closest to THOR’s target-hardware thesis.Mecka AIUnited States and Canada$60mMecka AI, $60m — Human-motion and task-specific training data for robotics labs, raised as $25m closed in November 2025 and a $35m follow-on. Body sensors and iPhones capture people performing physical tasks in real environments.MicroagiGermany$55mMicroagi, $55m — Licensed training data for frontier labs, and robots adapted for industrial customers, on a $55m seed in July led by Hummingbird with Northzone, LocalGlobe, Village Global and redalpine. It records experienced workers, which makes it a data vendor and a deployment business at once.LuelUnited States$31.2mLuel, $31.2m — Custom, rights-cleared multimodal datasets for robotics and other AI work, on a $31.2m seed in May backed by Lightspeed and General Catalyst. A global contributor network collects video and sensor data, with quality checks and provenance documented, across more than robotics.ConfigSouth Korea and United States$27mConfig, $27m — Bimanual manipulation datasets, with tools and models for training and deployment, on a $27m seed in May led by Samsung Venture Investment. It turns human and robot demonstrations into task-specific robot policies.Capital is flowing into robotics dataReported financing totals for the vendors shown, in US dollars. These companies span data collection, simulationand services for robotics developers.

Click any part for the detail

SourcesMecka AI15

Capital is flowing into companies that produce robotics data. THOR’s ambition is to widen who can participate in producing it.

  • We will open production to contributors wherever they have the skills to deliver. Contributors build a record of reliable work, while published datasets and tools give researchers a reason to participate. Each new participant can add expertise or capacity that the network’s customers can draw on.
  • The network brings people together around problems worth solving. Industry brings a task and a reason to pay. Researchers develop and evaluate approaches, entrepreneurs turn those advances into products, and investors fund the work needed to bring them to market. Open outputs let other teams contribute without rebuilding every component.

Fig 06

THOR will bring researchers, investors and entrepreneurs to one problemSiloed developmentOrganised around the problemCompanyResearchCapitalEngineeringCompanyResearchCapitalEngineeringCompanyResearchCapitalEngineeringIndustry names the problema need with a buyer behind itResearchers, investors, entrepreneurswork the problem togetherDeployment in real plantsperformance measured in the fieldThree roadmaps, three budgetsResults decide what gets funded nextSiloed development — Each company carries its own research, its own capital and its own engineering, and funds all three from one balance sheet. Progress accumulates inside the organisation that paid for it.Organised around the problem — Industry names a need with a buyer behind it. Researchers, investors and entrepreneurs work it together, each funded for the part they are best at. Deployment measures what happened, and that result feeds back into what gets researched and what gets financed next.SiloedOrganised around the problemOrganisationEach company pursues its own roadmapand builds within its boundariesIndustry defines the problem; participants coordinatearound common goals and measurable resultsOrganisation — A company can only work the problems inside its own walls, and it has to fund all of them. An industry that names the problem in the open lets several teams work it at once, each funded for the part it is best at.OutputsProgress accumulates withinindividual organisationsOpen hardware, open-weight models, datasets andevaluation tools become foundations others build onOutputs — Progress that stays inside one organisation has to be rebuilt by everyone else. Published hardware, weights, datasets and evaluation tools are the version that compounds, because the next team starts where the last one finished.DeploymentEach company develops its ownroute to customersEntrepreneurs turn shared advances into products,with industry partners testing them in practiceDeployment — Every siloed company pays to find its own customers. When the industry partner is in the loop from the start, the test is the deployment, and what it shows decides what gets developed and financed next.Deployment results feed back into research and into capital, so real-world performance decides what deserves furthersupport.

Click any part for the detail

Commercial demand should sustain a growing body of open robotics work.

Supporting research (2)
  1. 38

    Hugging Face LeRobot

    Hugging Face started LeRobot in 2024, bringing together open robotics models, datasets and tools. Its April 2025 acquisition of Pollen Robotics extended that effort into open hardware. Reachy Mini is a concrete example of the resulting open robotics platform.

  2. 15

    Mecka AI

    Founded in New York in 2025, Mecka raised $60m across a $25m Series A closed in November 2025 and a $35m follow-on, led by Framework with Menlo Ventures, SV Angel and Kindred; the valuation was not disclosed. It projects a $100m annual run rate off already-signed contracts, which is a projection rather than recognised revenue, with the customers unnamed. Collection runs through custom body sensors and iPhones rather than teleoperation rigs. Sequoia is reported to be leading a round valuing it at about $500m.

Section 06: Robotics needs a bridge no government will build.

Fig 07

Chinese manufacturers led humanoid volumes in 2025.ManufacturerUnits in 2025UnitreeChina>5,500Unitree, >5,500 units — Unitree discloses more than 5,500 customer deliveries of humanoid robots in 2025, excluding wheeled models, and is targeting 20,000 in 2026.AgiBotChina5,168AgiBot, 5,168 units — Omdia estimates 5,168 units for AgiBot in 2025. Its portfolio includes wheeled models, so the total is not measured the same way as Unitree’s.UBTECHChina1,079UBTECH, 1,079 units — UBTECH’s results disclose 1,079 full-size humanoids sold in 2025, with related products and services generating about RMB820m. An earlier Omdia estimate put shipments near 1,000.Figure AIUnited States≈150Figure AI, ≈150 units — Omdia estimates about 150 units for Figure AI in 2025, at a $39bn valuation.AgilityUnited States≈150Agility, ≈150 units — Omdia estimates about 150 Digit units for Agility Robotics in 2025.TeslaUnited States≈150Tesla, ≈150 units — Omdia estimates about 150 Optimus units for Tesla in 2025. Tesla has not disclosed customer sales of Optimus.ChinaUnited StatesEstimate, not reportedEvery bar counts units in 2025 on one scale. A solid bar is company-reported delivery or sales; an outlined bar is anOmdia estimate, and the two are not the same measure. Interact Analysis puts Chinese vendors above 90% of globalproduction, which passed 20,000 units, and about a tenth of what was produced went into real-world use.

Click any part for the detail

SourcesInteract Analysis22

  • A neutral bridge between China’s manufacturing strength and the world’s research is what THOR exists to build. Researchers should be able to access affordable hardware and commission the data they need, while manufacturers can draw on advances in open models and software.
  • Our relationships across Western and Chinese industry and research networks give us a starting point. A lab in Seoul should be able to commission data from a team in Shenzhen, with clear requirements, quality standards and rights to use the results.
  • Our prediction is that China’s robotics companies will be more selective about opening their models than its LLM developers have been. For a robot manufacturer, the ability to perform valuable work is a core competitive advantage. Sharing models can attract developers, while retaining commercial rights helps protect that advantage. Galaxea’s G0.541 illustrates this approach: its public release permits non-commercial use, while commercial deployment requires a separate licence.
  • Researchers building their own models will still need access to training data. An independent data layer gives them a way to develop capabilities alongside both open and proprietary systems.
  • Export controls and political tensions29 can make collaboration across borders harder. That makes a network with multiple sources of hardware, research and data more important. We believe robotics should remain a global effort, even as governments pursue different strategies.
Supporting research (4)
  1. 22

    Humanoid production and unit volumes in 2025

    Interact Analysis puts global humanoid production above 20,000 units in 2025, a tenfold rise on 2024, with Chinese vendors above 90% of it, and only about a tenth of what was produced deployed in real-world applications. Company disclosures put Unitree above 5,500 customer deliveries, excluding wheeled models, and UBTECH at 1,079 full-size humanoids sold. Omdia estimates AgiBot at 5,168 and Tesla, Figure AI and Agility Robotics at about 150 each. Production, shipments, deliveries and sales are four different measures.

  2. 23

    Chinese government robot-data centres

    Chinese local governments funded a large number of humanoid training centres to solve the robot-data shortage. British firm Interact Analysis reported more than 40 government-owned robot data-collection centres were slated to open by December 2025, with over 20 already operating.

  3. 41

    Galaxea G0.5 licensing

    Galaxea releases G0.5 model code, weights and accompanying materials under its Community License for non-commercial purposes, including research, education and evaluation. Commercial use, including production deployment and productisation, requires a separate commercial licence. This is an example of selective commercial access, not evidence that all Chinese robotics companies follow the same approach.

  4. 29

    US chip export policy, 2025

    Policy toward China reversed repeatedly: the H20 was banned in April, allowed again in July, and in December 2025 the administration moved to allow H200 exports with a revenue cut to the government. It illustrates how volatile decoupling is.

Section 07: Open hardware brings more laboratories into robotics.

  • Reproducibility starts with what a lab can actually build: affordable arms, hands and capture kits make participation possible. SO-10130, LEAP Hand31 and other open designs37 give teams published hardware and tools they can reproduce and modify.
  • Standard hardware makes data easier to share. When labs collect demonstrations on common platforms, more of that work can be reused. DROID12, for example, used the same hardware setup across 13 institutions. Calibration and preparation still matter, and data may need adaptation between robot bodies, but standardisation reduces the work each lab has to repeat.
  • The open designs below give laboratories more ways to build and maintain their equipment, reducing dependence on a single supplier. Affordable arms and hands widen access to research; the demonstrations needed to train them still have to be produced.

Fig 08

Open robotics is reproducible in any country,so no single supplier becomes indispensable.Open projectWhat is publishedCost or licenceAmazing HandPollen Robotics, FranceBill of materials, printable designs,assembly instructions and demo programsunder €200Amazing Hand, Pollen Robotics, France — Pollen Robotics grew out of Inria research on Poppy, an early open-source humanoid, and is now part of Hugging Face. Amazing Hand publishes its bill of materials, printable designs, assembly instructions and demo programs, with eight degrees of freedom for a published parts cost below €200. It is an experimental platform rather than a substitute for every research hand.ORCA HandETH Zurich, SwitzerlandHardware files and control softwarefor a 17-degree-of-freedom hand$2,000 in partsORCA Hand, ETH Zurich, Switzerland — The ORCA Hand came out of ETH Zurich’s Soft Robotics Lab, a 2025 research hand with 17 degrees of freedom and a reported $2,000 bill of materials. The hardware files and control software are public, though the v2 touch-sensing module remains proprietary, so the whole product stack is not open.OpenArmEnactic, JapanCAD, firmware, control code andsimulation tools, under CERN-OHL-S-2.0CERN-OHL-SOpenArm, Enactic, Japan — Enactic builds an openly reproducible arm platform for physical AI research, teleoperation and evaluation. CAD, firmware, control code and simulation tools are published, with the hardware repository under CERN-OHL-S-2.0, and kits and assembled units are sold alongside it.LEAP HandCarnegie Mellon, United StatesOpen build resources and software,about four hours to assemble$2,000 to buildLEAP Hand, Carnegie Mellon, United States — LEAP is a robot-learning research project rather than a company. The original hand publishes a $2,000 build cost from readily available components and roughly four hours of assembly. Later LEAP designs differ substantially, so the version matters.Solo and TriFingerMax Planck, LAAS-CNRS and NYUDrawings, electronics and software formodular actuators, under BSD 3-ClauseBSD 3-ClauseSolo and TriFinger, Max Planck, LAAS-CNRS and NYU — A collaboration involving Max Planck, LAAS-CNRS and NYU, designed so that other laboratories can reproduce low-cost research robots. Mechanical drawings, electronics and software for modular actuators, legged robots and manipulation platforms are published under BSD 3-Clause.Berkeley Humanoid LiteUC Berkeley, United StatesSchematics, bill of materials, firmware,control software and the training codeunder $5,000Berkeley Humanoid Lite, UC Berkeley, United States — A 3D-printed humanoid built around modular printed gearboxes and parts anyone can order, at about $4,300 in the United States and $3,200 in China. Schematics, bill of materials, firmware, control software, simulation environments and reinforcement-learning code are all published.Six projects, five countries, one reproducible stackPublished build costs and licences come from each project. Every design here can be rebuilt by a laboratory that hadno part in making it.

Click any part for the detail

Sourcesopen hardware37LEAP Hand31

THOR will pay for more open hardware to be published.

Supporting research (4)
  1. 30

    Hugging Face SO-101 arm

    An open 3D-printable arm that can be built for as little as $130 in parts, with off-the-shelf kits from about $220. It is natively supported by LeRobot.

  2. 31

    LEAP Hand

    A 16-DoF anthropomorphic open-source hand assembled in about 4 hours for $2,000: one-eighth the cost of the Allegro Hand at $15,000, and a fraction of the Shadow Hand at about $100,000.

  3. 37

    Open robot hardware, published designs

    Pollen Robotics grew out of Inria work on Poppy and is now part of Hugging Face; its Amazing Hand publishes a bill of materials, printable designs, assembly instructions and demo programs for a parts cost below €200. The ORCA Hand, from ETH Zurich’s Soft Robotics Lab, is a 17-degree-of-freedom platform with a reported $2,000 bill of materials, though its v2 touch-sensing module stays proprietary. Enactic’s OpenArm publishes CAD, firmware, control code and simulation tools with the hardware repository under CERN-OHL-S-2.0. The Open Dynamic Robot Initiative, a Max Planck, LAAS-CNRS and NYU collaboration, publishes drawings, electronics and software for modular actuators and legged platforms under BSD 3-Clause. Berkeley Humanoid Lite publishes schematics, bill of materials, firmware, control software and training code for a 3D-printed humanoid costing about $4,300 in the United States.

  4. 12

    DROID

    An in-the-wild robot manipulation dataset with 76,000 demonstration trajectories, 350 hours of interaction data, 564 scenes and 86 tasks. Its 13 participating institutions used the same robot hardware setup to streamline collection and improve portability, and the project publishes a guide for reproducing that setup.

Section 08: THOR Group will build the infrastructure for physical intelligence.

THOR Group will bring together businesses with distinct customers and capabilities, connected through SN112. Their work spans data production, research, hardware and real-world deployment.

  • THOR Data organises annotation and commissioned collection around researchers’ needs. THOR Simulation extends that work by reconstructing captured scenes and generating variations for training and evaluation.
  • THOR Research backs university teams and independent laboratories publishing usable model weights, training code and evaluation tools.
  • THOR China develops manufacturing partnerships and commercial relationships, connecting the group’s work with customers and suppliers in China’s robotics industry. It is led by Wenji Cheng, who brings experience from Sequoia Capital China, now HongShan.
  • THOR Deployment works with industry to integrate models and hardware into real operations. THOR Hardware supports that work with capture kits and reproducible designs that other teams can build on.

Fig 09

THOR’s companies work in synergy,all feeding one network that supports researchTHOR Group (SN112)the SubnetSupplyDataTHOR Datademonstrations, operationaldatasets and simulationCredibilityResearchTHOR Researchopen-weight models, researchpartnerships and evaluationMarketChinaTHOR Chinacustomer development andmanufacturing partnershipsRecurringDeploymentTHOR Deploymenthardware, systems integrationand real-world operationTHOR Simulationone capture, many episodesTHOR Hardwarecapture kits, installed baseTHOR Group (SN112) — The group holds every unit outright. Each one sells into a market that already pays for what it does, and each returns data to the same network, so revenue earned in one unit raises the supply the others train on.THOR Data — Customer orders become human demonstration data captured on the target hardware, and simulation data generated from it.THOR Simulation — A scene captured once is rebuilt in software, then lighting, objects, positions and physics vary and every variation is a new episode. A simulated scene already knows its own configuration, so nothing is left to annotate frame by frame.THOR Research — Support for early-stage research institutions doing frontier robotics work, on acceptance tests anyone can reproduce in an afternoon, and new robotics models come out of it.THOR China — The unit that commercialises inside China, led by Wenji Cheng, so that what THOR builds reaches the largest hardware manufacturing base in the world.THOR Deployment — Partnerships with industry to deploy manufacturing built on the models, and incentives that get more open hardware built, so the capture kits those deployments run on keep multiplying.THOR Hardware — Open capture kits built on the stack every lab already runs, and incentives paid to anyone who publishes more of them, so each deployment grows the installed base and returns new data supply.Each unit is measured against its nearest comparables in its own market, listed in its detail card.

Click any part for the detail

  • Pylon Partners’ experience operating Bittensor subnets and commercialising open-source technology gives THOR a starting point. Our contributors and advisors bring experience in annotation delivery, robotics data operations and hardware deployment. Together, they connect the requirements of research with the practical work of fulfilling them.
  • We are building these businesses to strengthen one another. Research identifies gaps in capability; data work addresses them; hardware and deployment test the results in practice. Those results guide the next round of collection. A customer who begins with a data order may later need capture hardware or help deploying a model. Serving those needs gives THOR more ways to earn from the same relationship, while experience delivering the work improves the next project. We will expand from annotation as customer demand and demonstrated performance justify it.
Explore THOR’s expansion from data to deployment

Fig 10

THOR starts with data and expands across roboticsTHOR1231DataCollect, annotate and verify what labs need2Research & ModelsHelp researchers train and improve open models3Hardware & DeploymentSupport open hardware and bring robotsinto industryReal-world use generates new data1. Data — Demonstrations captured on the target hardware, annotated and checked against what the buyer asked for. It sits in the middle because it is the input nobody can scrape, and the one every layer outside it depends on.2. Research & Models — Open-weight models, research partnerships and evaluation anyone can reproduce. This layer is what the data is collected for, and what the labs outside THOR build on.3. Hardware & Deployment — Open designs that get built, and industry partners running them in real plants. It is the outermost layer because it is the last to arrive and the first to prove whether any of the rest of it worked.Real-world use generates new data — Every deployment runs on a model, and every model that runs in a real plant produces the thing the middle of this diagram is short of. Expanding outward is what keeps the centre supplied, which is why the layers are drawn around each other rather than beside each other.

Click any part for the detail

  • As robots enter more workplaces, we expect demand to grow for the data, models and integration that make them productive.
  • Each business must earn its place in its own market. We will build the group’s advantage by sharing customer relationships, technical knowledge and data across those businesses. Our intended model is for the commercial value they create to accrue to SN112, linking the network’s contributors and token holders to the group’s long-term success.

Fig 11

THOR’s businesses are intended to deliver value to SN112 holdersRevenueData Marketplaceverified demonstrations,sold at a stated priceCredibilityIncentivise Researchemissions paid to the workthat moves the fieldInstalled baseCommercialise Hardwareopen designs, built anddeployed in real plantsCommercial valueSN112the SubnetData Marketplace — Buyers name the footage, the acceptance rubric and the dollar rate, and the network delivers against it. Every order leaves behind a data asset that did not exist before it was placed.Incentivise Research — Emissions pay for open-weight models, evaluation anyone can reproduce, and the early-stage institutions doing frontier robotics. What that work publishes is what the rest of the group builds on.Commercialise Hardware — Open designs get built, and industry partners put them into real plants. Each installation grows the base the group collects from, and pays for the next set of designs to be published.SN112, the Subnet — Our intended model is for commercial value generated across THOR’s businesses to accrue to SN112. The thesis links contributors and token holders to the group’s long-term success, while each business strengthens the network through data, expertise and customer relationships.

Click any part for the detail

THOR will build businesses that strengthen the network as they grow.

THOR’s approach at a glance

Fig 12

Robotics needs usable data and shared infrastructureTHOR coordinates the work between research and real-world deployment.The bottleneckTHOR’s responseData that cannot train the robotRaw footage lacks reliable action labels and hardwarecompatibility.Commission what labs needResearchers define the task and the hardware, and contributorscollect and annotate against those requirements.Data that cannot train the robot → Commission what labs need — Hours of video is not the same thing as training data. Commissioning starts from the requirement, so what comes back is already calibrated, labelled and matched to the robot it is meant to train.Progress locked inside companiesPrivate datasets and closed models limit what other teams canbuild on.Build shared foundationsSupport open datasets, open model weights and reproduciblehardware.Progress locked inside companies → Build shared foundations — A closed policy trains one company’s robots. An open one trains the field’s. Shared datasets, published weights and hardware anyone can rebuild are what let the next team start from where the last one stopped.Research disconnected from deploymentResearchers, investors and entrepreneurs pursue separatepriorities.Organise around industry problemsMatch research and funding to customer needs, then measure theresults in deployment.Research disconnected from deployment → Organise around industry problems — Inside one company those three sit on a single cap table competing for the same budget, and the industry problem comes third. Organised around the problem, the benchmark, the return and the contract are the same piece of work.Capabilities divided across bordersManufacturing expertise and frontier research remainfragmented.Connect global capabilitiesBring international research together with manufacturingpartners and customers.Capabilities divided across borders → Connect global capabilities — The research strength and the manufacturing strength are not in the same place, and no government is going to join them. A network that pays for verified work wherever it happens is the bridge neither side will build.

Click any part for the detail

Our conviction

The world should help build what comes next.

Physical intelligence will shape manufacturing, healthcare, science and daily life. Who teaches machines to act, and who gets to build with that knowledge, will matter to everyone.

I started THOR Group with our partners because we believe that globalisation and collaboration can accelerate robotics access for all. Our conviction rests on three principles.

Open participation: capable contributors should be able to help build robotics wherever they are, with their work recognised and rewarded.

Verified usefulness: clear requirements and measured performance should determine what earns trust, payment and further investment.

Shared progress: published datasets, usable models and reproducible hardware should give each new team a stronger starting point.

Bring the problems you need robots to solve. Contribute your data, tools or expertise. Help build capabilities that others can use and improve.

Humanity is about to give intelligence a body. THOR will help build the open foundation on which it learns to move.

Jack Ai-Leung
On behalf of THOR Group

Build physical intelligence in the open.

Explore the networkRead the docs