Chris Feng, co-founder and CEO of Axis Robotics, raised a $12 million seed round for a system that turns browser users into distributed operators generating robot-training data. Axis announced the financing on July 27th and reiterated it in a post on X. Hack VC led the round, with Nomad Capital, Pi Network Ventures and 10K Ventures participating alongside angel investors.
https://x.com/axisrobotics/status/2081711331791827387 Feng previously served as chief operating officer at Chainbase, according to a public professional profile. His bet at Axis Robotics applies a contributor-network playbook to a problem robotics labs have struggled to solve with conventional data operations. "Physical AI demands billions of human-physical interaction motion trajectories," Feng said in the sponsored funding announcement.
Founded in 2025 and headquartered in Berkeley, California, according to its LinkedIn profile, Axis Robotics is building infrastructure around the premise that robot models need far more varied demonstrations than labs can gather with physical machines and centralized operators. The round gives Feng capital to test whether a global pool of lightly equipped contributors can produce data that survives the jump from simulation to real hardware.
A web browser as a teleoperation lab
Axis Robotics lets contributors control simulated robot arms through a browser. A completed session produces a trajectory containing actions, robot joint states, object positions and task metadata. Axis Robotics says its pipeline then checks whether the task succeeded, filters and smooths the trajectory, and uses visual and physics-based augmentation to prepare the data for vision-language-action models and imitation-learning policies.
The product stack combines a procedural task-generation engine, browser-based simulation, automated quality controls and downstream model evaluation. Randomizing layouts, camera positions, objects, lighting and robot configurations is central to Feng's pitch: increasing variety should make a policy less dependent on the narrow conditions present in a single lab or fixed benchmark.
Axis Robotics has also built an on-chain record around the collection process. Its documentation says each accepted trajectory receives a Data ID recorded on Base, creating an ownership and provenance trail. That record can show where a trajectory came from and who contributed it. It does not establish whether the demonstration is useful. Axis Robotics still depends on success checks, filtering, training runs and real-robot evaluations to determine quality.
The company extended the system in July with Axis V2, which focuses on collecting corrections against a model's current failures. Contributors encounter states produced by a policy, intervene when it goes wrong and feed the corrected trajectory into subsequent training. That approach could make the contributor network more valuable over time because collection targets model weaknesses instead of continually adding generic demonstrations.
The research paper narrows the claims
A technical paper submitted to arXiv on July 23rd provides the clearest public test of the system. The researchers describe a dataset containing 207 tasks and more than 50,000 trajectories, with automated task generation, success checking, quality filtering, trajectory smoothing and visual and physics-based augmentation.
The paper reports that continual pretraining on the AXIS dataset improved the overall success rate of the pi0.5 vision-language-action policy by 5.8% and outperformed a version pretrained on RoboCasa365 by 37.3%. The largest gains appeared under changes to layouts, camera inputs, sensor noise and robot poses, according to the abstract. Those are precisely the conditions a diversity-focused data engine should address.
Axis Robotics' July 27th sponsored announcement cited different benchmark figures: a 4.9-point improvement and a 31.3-point advantage over the volume-matched RoboCasa365 baseline. The paper and announcement may be describing different dataset versions or calculations, but the differing figures make the evaluation protocol and comparison set essential context when assessing the results.
The company's scale metrics also describe different parts of the system. Axis Robotics' website claims a network of more than 100,000 contributors, while its documentation lists more than 200,000 verified trajectories. The paper evaluates a curated dataset of more than 50,000 trajectories. Network registrations, accepted production trajectories and research-set trajectories are separate measures; combining them would exaggerate the amount of data tested in the published benchmark.
The investor list explains the distribution strategy
The backers point to the second half of Feng's plan. Hack VC describes itself as an investor in Web3 and AI. Nomad Capital and 10K Ventures focus on crypto companies, while Pi Network Ventures invests with the stated goal of expanding real-world uses for Pi Network. Their involvement fits Axis Robotics' use of contributor incentives and on-chain attribution.
That makes the financing a distribution bet as much as a robotics research bet. Axis Robotics needs enough contributors to cover a wide range of tasks and edge cases, then enough automated controls to keep low-quality submissions from overwhelming the useful data. The company said the seed capital will expand its procedural generation technology and contributor network.
Axis Robotics says it commercializes the output as customized "Task Packages" for robot hardware manufacturers, model developers and industrial users. Each package is organized around a deployment scenario, required skills, randomized conditions and trajectory volume. The model gives robotics companies a way to commission data around specific failure modes without building a large teleoperation facility of their own.
The $12 million buys Feng time to prove that distributed simulation produces data customers will purchase repeatedly. Axis Robotics has shown benchmark improvements and described a path from browser demonstrations to physical execution. The commercial test is whether those trajectories reduce failures across customer hardware, environments and tasks enough to justify recurring spending. If they do, Axis Robotics can sell the data-production layer to robotics companies that cannot afford to build one at comparable scale.