Axis Robotics, a startup developing data infrastructure for Physical AI, has raised $12 million in a seed round led by Hack VC, with participation from Nomad Capital, Pi Network Ventures, 10K Ventures, and several angel investors. Announced on July 27, the investment comes amid growing demand for robot training data as robotics companies expand deployments beyond testing environments.
Axis stated it will use the capital to expand its data engine for robot training, aiming to build a pipeline for continuous data generation and improvement for Physical AI systems.
We’re thrilled to announce a $12M Seed round, led by @hack_vc, with participation from @NomadCapital_io , @PiCoreTeam Ventures , @10kventure and top angel investors.
Physical AI has a data problem. Models need more than static datasets—they need diverse data that evolves with… pic.twitter.com/byx4JSC7fn
— Axis Robotics (@axisrobotics) July 27, 2026
A $12M Bet on Physical AI Data
The seed round places Axis among the startups building data infrastructure for Physical AI, rather than developing robots or foundation models. Led by Hack VC with participation from Nomad Capital, Pi Network Ventures, and 10K Ventures, the deal reflects a trend of investors beginning to view robot training data as an infrastructure layer capable of scaling alongside the robotics market.
This thesis stems from a common industry challenge: Physical AI systems cannot rely solely on datasets collected just once. As robots are deployed in real-world environments, models must continuously ingest additional data from new scenarios, detect errors, and update policies to improve performance over time.
Instead of competing on hardware or foundation models, Axis aims to build the infrastructure to generate, validate, and update data for the robot training process, targeting Physical AI development teams in need of data sources that can scale with their deployments.
Inside Axis’s Data Engine
Axis’s core product is a closed-loop data engine for robot training, combining large-scale simulation, real-world egocentric data, and a human-in-the-loop post-training process.
Axis’s data engine combines three layers of data. The first is large-scale simulation to generate robot trajectories across various environments, tasks, and robot embodiments. Next is egocentric data collected from the robot’s perspective in real-world environments. Finally, the company utilizes a human-in-the-loop process to review, correct errors, and improve policies during the post-training phase.
In its year-end roadmap, Axis plans to deploy human-gated DAgger — a variant of the imitation learning method that only requires human intervention when the robot makes incorrect decisions or needs correction. The company expects this approach to help reduce the cost of generating post-training data while maintaining the quality of data for training.
According to Axis, the company’s system has processed over 200,000 verified trajectories. Previous campaigns also recorded 10,000+ valid trajectories in 3 days and 100,000 trajectories in 5 days.
The Bottleneck Holding Back Robots
Unlike language foundation models, which are trained on massive amounts of internet data, Physical AI must learn from real-world interactions — where every action is tied to objects, spaces, physical forces, and various environmental conditions.
This makes robot training data significantly harder to scale. Data is often fragmented by robot type, task, hardware, and deployment environment, while a policy that works well on one robot may not necessarily transfer to another. The gap between simulation and real-world operating conditions also continues to be a major barrier to commercial-scale robot deployment.
Consequently, many robotics companies are shifting their attention to platforms capable of continuously generating and updating data, rather than merely scaling models or hardware.
What’s Next for Axis
Following the seed round, Axis will focus on expanding both its product capabilities and operational scale. In the coming months, the company expects to deploy an egocentric data pipeline in September, expand simulation to more robot embodiments and atomic capabilities in October, and launch a large-scale post-training dataset based on human-gated DAgger by the end of the year. According to Axis, the company has collected “tens of thousands of hours” of egocentric data and is co-developing product requirements with several frontier labs.
Alongside product expansion, Axis also aims to scale its contributor network. The company stated it currently has over 100,000 contributors and aims to expand into Latin America and Eastern Europe, while increasing daily active users to 10,000. Operationally, Axis aims to generate over 500 hours of egocentric data and 50 hours of simulation data daily, while also developing the capacity to generate corrective post-training data.
On the commercial front, Axis aims to complete two to three paid pilots before the end of the year and become a preferred vendor for foundation model development companies in Q1 of next year. In the long term, the company wants to integrate its data engine directly into the training and deployment workflows of robot developers, AI model developers, and industrial operators.
Although the roadmap is fairly well-defined, Axis still needs to prove that data generated from crowdsourcing combined with simulation can improve performance during real-world robot deployment, rather than just scaling the dataset. This outcome will determine whether the company’s data infrastructure model can become a critical infrastructure layer for Physical AI as the industry transitions from initial experiments to commercial-scale deployment.
