Axis Robotics has raised $12 million in seed funding to scale a data-collection platform that recruits remote contributors to generate training material for robots, targeting one of physical AI’s most persistent bottlenecks: the shortage of diverse, usable motion data.
Hack VC led the round, with Nomad Capital, Pi Network Ventures and 10K Ventures participating, Axis said. The company plans to use the capital to expand its human-in-the-loop Compounding Data Engine, a system intended to create and refine datasets used to train robotic models for manipulation and other physical tasks.
Axis is positioning the platform as an alternative to the slower, more expensive process of collecting robot demonstrations in controlled laboratories. Its approach combines simulated tasks, remote teleoperation and mobile real-world capture, then feeds the resulting data through a processing pipeline designed to prepare it for model training.
The funding places Axis among a growing group of robotics infrastructure companies seeking to turn data generation into a scalable service. Better foundation models alone may not solve the problem of getting robots to work reliably across unfamiliar rooms, object layouts, camera conditions and hardware configurations; they need demonstrations that capture those variations.
a remote workforce for robot demonstrations
Axis said its contributor network has more than 100,000 active participants, each submitting data an average of three to four times per day. The company said that network can generate more than 1,200 hours of simulation data and over 20,000 hours of real-world, ego-centric data each month.
Those figures, if sustained at scale, would give robotics developers access to data volumes that are difficult for individual hardware companies to produce internally. Conventional data collection often requires dedicated robots, lab space, operators and safety oversight. Axis instead aims to distribute parts of that work through browser-based and mobile tools.
Its browser-based simulation teleoperation interface allows contributors to remotely control robots in virtual settings and record motion trajectories, or paths describing how a robot should move during a task. Axis said this system delivers ten times the throughput of lab-based collection, though it did not provide the underlying comparison methodology.
The company’s mobile application is intended for real-world capture through real-time hand-pose tracking. Rather than asking contributors to operate specialized robotics equipment, the app is designed to record human movement that can be transformed into demonstrations for robotic systems.
That setup could reduce the logistical friction of collecting examples of everyday tasks such as picking up items, opening containers or arranging objects. The usefulness of such data depends heavily on how accurately human movements can be mapped to robot actions, particularly where a robot’s gripper, joints or camera arrangement differs from the human demonstrator’s body.
building data for different robots and environments
Axis said it is trying to address three connected challenges: scarcity of robotics data, the difficulty of generalizing from training conditions to new environments, and variation between robot hardware platforms.
A robot policy trained on demonstrations from one room or machine may fail when lighting changes, objects move, sensors become noisy or a different robot arm performs the same task. Axis says its engine is designed to create data across those types of variation rather than merely increase the number of examples of a single task.
The system’s processing pipeline cleans trajectories, applies domain randomization and adds dense language annotations, according to the company. Domain randomization changes aspects of a training environment—such as object positions, visual appearance or lighting—so that models encounter more variation before operating outside a simulation.
Axis said the pipeline produces multimodal datasets, combining different forms of information such as visual inputs, motion and language. It also claimed more than a tenfold improvement in data quality, without defining the measurement used for that claim.
The workflow includes human-gated DAgger intervention loops. DAgger, short for Dataset Aggregation, is a training technique that adds corrective examples when a model reaches situations where its prior training data does not adequately guide it. In practice, this can help focus human input on failures or difficult edge cases instead of requiring people to collect every demonstration from scratch.
early benchmark results and commercial customers
Axis has released Sim Dataset V1 and said it tested the dataset on LIBERO-Plus, a benchmark used to evaluate robot manipulation tasks. According to Axis, pretraining the π0.5 model on its diversified dataset increased overall success by 4.9 percentage points.
The company also said the same model outperformed a volume-matched RoboCasa365 baseline by 31.3 points. Axis attributed the gains to stronger performance across layout changes, sensor noise and changes in robot pose.
Benchmark improvements offer an early indication that a dataset can help a model handle variation, but deployment conditions remain harder than structured evaluations. Industrial and consumer robotics companies must also contend with hardware reliability, safety constraints, latency and the cost of integrating models into specific machines.
Axis plans to sell customized “Task Packages” to robotics hardware manufacturers, physical AI model developers and industrial automation companies. The packages appear intended to provide tailored datasets for particular hardware, tasks or operating environments rather than a single general-purpose corpus.
The company listed Booster Robotics, Manycore Tech, Feagine Robotics, Dexmal, Lotus Car, Geely Auto and SomaStacks among its early partnerships. Axis did not describe the commercial scope, duration or technical details of those arrangements.
data infrastructure becomes a robotics battleground
The company’s strategy rests on a practical view of the robotics market: data collection may become as consequential as model architecture for teams trying to deploy machines beyond narrowly scripted tasks. A large contributor network gives Axis a potential mechanism to capture more task variety than a centralized lab operation, while simulation can generate low-cost training examples before real-world validation.
Its ability to convert that supply into consistently useful training material will determine whether the model works economically. More data can add noise as easily as it adds coverage, especially when contributors use different devices, operate in different environments or follow task instructions unevenly.
Axis said its team includes AI and robotics researchers from the University of California, Berkeley, Carnegie Mellon University, Georgia Tech, Nanyang Technological University and Shanghai Jiao Tong University. It added that some team members previously helped scale consumer products to more than 30 million global users.
With the new funding, Axis will be judged less by contributor counts than by whether its task packages help customers shorten the path from model training to reliable robot behavior on real hardware.
Curious how AI transforms trading like robotics? Explore automated strategies with AI copy trading on Toobit.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
