Axis Robotics releases major open Franka arm simulation dataset, raises $12M
Axis Robotics released Axis Sim Dataset V1, an open-source Franka arm manipulation dataset with full training code and benchmarks. V1 contains 50,000+ human-teleoperated simulation trajectories across 207 tasks and 60,000+ scene variants on a simulated Franka Research 3 arm. It surpassed 160,000 downloads on Hugging Face, the highest among open-source Franka manipulation simulation datasets. Benchmarks show continual pretraining on V1 lifts π0.5 and outperforms a volume-matched RoboCasa baseline, with results publicly verifiable.
Axis is building a compounding data engine for Physical AI that integrates large-scale simulation, egocentric real-world capture, humanoid loco-manipulation, and human-gated DAgger post-training. The company raised $12 million in seed funding led by Hack VC, with participation from Nomad Capital, Pi Network Ventures, 10K Ventures, and angel investors.
Thesis: Scale and Diversity Over “Clean Data Only”
Axis posits that data quality emerges from distributional scale and diversity. When a large, diverse crowd produces noisy, suboptimal trajectories with uncorrelated errors, averaging during training yields robust policies. V1 tests this: trajectories span pick-and-place, stacking, pouring, articulated-object manipulation, and tool use, collected via Axis Hub’s browser-based teleoperation by distributed contributors. Collaborators include UC Berkeley, Johns Hopkins, and the University of Michigan.
Benchmark Results
On LIBERO-Plus, continual pretraining on V1 increases π0.5 success from 83.9% to 88.8% and exceeds a volume-matched RoboCasa365 baseline by 37.3%. Performance scales monotonically when using 25% to 100% of V1 with no saturation observed. Largest gains appear under camera, sensor-noise, and layout perturbations, which V1 explicitly randomizes.
V2 is in progress: 1.2 million trajectories across 1,200 tasks, targeting cross-embodiment generalization and evaluations across multiple VLA models to validate robustness from suboptimal simulation data.
Compounding Data Engine
Axis collects iteratively based on model performance and failure cases. Four scaled data lines:
- Simulation: 200,000+ contributors on Axis Hub, a top-3 dApp on Base, producing 4.7 million+ trajectories across 13 embodiments.
- Egocentric: 1,000+ full-time QC-trained collectors capturing first-person data across 14 industries: 200,000+ hours banked, growing by 4,000+ hours daily, with Vicon-verified hand pose.
- Loco-manipulation: 500+ hours on real humanoids, including Unitree G1 and Booster T2, via hardware-agnostic teleoperation.
- Human-gated DAgger post-training: 500+ hours of human-in-the-loop correction focused on deployment edge cases.
All tasks and trajectories are recorded on-chain on Base for provenance, and contributors receive rewards tied to verified work quality.
From Open Data to Deployment
Axis partners with robot embodiment companies to build custom data pipelines and priors. With Booster Robotics, Axis created a digital twin of the real workspace, collected 42,000+ simulation episodes, and distilled a Booster-specific prior. Using 30 real-robot demos, that prior reached 87.5% success versus 37.5% for an out-of-the-box π0.5, matching π0.5 with half the real-world demos.
Additional partners: Feagine Robotics, Manycore Tech, Dexmal, Lotus Cars, Geely Auto. Axis supplies on-chain robotics networks including BitRobot on Solana and OpenRoboto on Bittensor.
Organization
Axis was founded by researchers from UC Berkeley, CMU, Georgia Tech, and SJTU, and serial founders with products scaled to 30 million+ users. Research advisor: Jiachen Li, Assistant Professor at Georgia Tech. Founder Chris Feng states that Physical AI requires a closed-loop engine that scales, diversifies, and compounds data based on model needs.
Paper: https://arxiv.org/abs/2607.21588
Project: https://axisaiorg.github.io/AXIS-V1/
Dataset: https://huggingface.co/datasets/axisrobotics/Franka-Dataset
Code: https://github.com/AxisAIOrg/Axis-V1-Training
