Axis Robotics unveils open-source Franka arm dataset for Physical AI

Axis Robotics has made Axis Sim Dataset V1, one of the largest open-source simulation datasets for Franka arm manipulation, publicly available along with the full dataset, training code, and benchmarks. V1 comprises over 50,000 human-teleoperated simulation trajectories spanning 207 manipulation tasks and more than 60,000 scene variants on a simulated Franka Research 3 arm.

The dataset has attracted more than 160,000 downloads, becoming the most downloaded open-source Franka simulation dataset on Hugging Face. In benchmark tests, continual pretraining on V1 improved π0.5 and outperformed a volume-matched RoboCasa baseline, with all results open and verifiable.

Axis Robotics is developing a compounding data engine for Physical AI, a vertically integrated system covering large-scale simulation, egocentric real-world capture, humanoid loco-manipulation, and human-gated DAgger post-training. The firm secured $12 million in seed funding, led by Hack VC, with Nomad Capital, Pi Network Ventures, 10K Ventures, and angel investors also participating.

A Bet Against “Clean Data Only”

A widely held belief in robotics is that demonstrations should be near-optimal, filtering down to expert trajectories and discarding noisy data. Axis’s thesis argues the opposite: data quality is defined at the distribution level, not per trajectory. When a large and diverse crowd produces noisy, suboptimal trajectories with uncorrelated errors, the noise averages out and a working policy emerges during training. Axis Sim Dataset V1 is a public test of that thesis. The trajectories cover pick-and-place, stacking, pouring, articulated-object manipulation, and tool use, collected via Axis’s browser-based teleoperation platform, Axis Hub, by a distributed crowd. The dataset was developed with researchers from UC Berkeley, Johns Hopkins, the University of Michigan, and others.

Results That Scale

On LIBERO-Plus, continual pretraining using V1 boosts π0.5 success from 83.9% to 88.8% and surpasses a volume-matched RoboCasa365 baseline by 37.3%. Performance steadily increases as pretraining data is scaled from 25% to 100% of the dataset, with no signs of saturation, indicating gains come from diversity and coverage. The largest improvements occur under camera, sensor-noise, and layout perturbations, the very axes Axis randomizes during generation. The team reports that V2 is already in progress, scaling to 1.2 million trajectories across 1,200 tasks, with cross-embodiment generalization and results across multiple VLA models suggesting suboptimal simulation data can train robust policies.

The Engine Behind the Dataset

The dataset is one product of a larger, continuously compounding data engine. Unlike traditional data vendors that collect to a fixed specification and stop, Axis uses model performance and failure cases to guide next collection, so each training round informs the next. The engine operates on a hybrid strategy across four data lines, all now at scale:

• Simulation: more than 200,000 distributed contributors on Axis Hub, a top-3 dApp on Base, generating over 4.7 million trajectories across 13 embodiments.
• Egocentric: a managed network of over 1,000 full-time, QC-trained collectors capturing first-person activity in real homes and businesses across 14 industries, with 200,000+ hours banked and growing by 4,000+ hours daily, verified by Vicon hand pose.
• Loco-manipulation: 500+ hours combining mobility and dexterity on real humanoids such as Unitree G1 and Booster T2 via hardware-agnostic teleoperation.
• Human-gated DAgger post-training: 500+ hours of human-in-the-loop correction targeting deployment edge cases.

Every task and trajectory is recorded on-chain on Base for provenance, and contributors are rewarded based on verified work quality.

From Open Data to Commercial Deployment

In addition to open-sourcing simulation data, Axis collaborates directly with robot embodiment companies to develop customized, embodiment-specific data pipelines and model priors. As Booster Robotics’ first simulation-data partner, Axis reconstructed Booster’s real workspace as a task-aligned digital twin, had distributed contributors collect over 42,000 simulation episodes, and distilled them into a Booster-specific model prior. With only 30 real-robot demonstrations, that prior achieved 87.5% success, compared to 37.5% for an off-the-shelf π0.5, matching π0.5 using half the real-world demonstrations. Other partners include embodiment companies (Feagine Robotics), model companies (Manycore Tech, Dexmal), and industrial automation firms (Lotus Cars, Geely Auto). Axis also supplies on-chain robotics networks such as BitRobot on Solana and OpenRoboto on Bittensor.

Redefining Physical AI’s Data Foundation

“The future of Physical AI isn’t a static dataset you download once,” said Chris Feng, founder of Axis Robotics. “It’s an engine that keeps producing the data the model needs next. Scale gets you broad coverage. Diversity keeps the noise unbiased. The closed loop turns every failure into progress. That’s what compounds.”

Axis was founded by researchers from UC Berkeley, CMU, Georgia Tech, and SJTU, as well as serial founders who have scaled consumer platforms to more than 30 million users. Its research is advised by Jiachen Li, assistant professor at Georgia Tech.

The paper is available at https://arxiv.org/abs/2607.21588.

The project page is at https://axisaiorg.github.io/AXIS-V1/.

The dataset is at https://huggingface.co/datasets/axisrobotics/Franka-Dataset.

The GitHub codebase is at https://github.com/AxisAIOrg/Axis-V1-Training.