Axis Robotics Releases Large Open-Source Franka Arm Simulation Dataset

Axis Robotics announced the public release of Axis Sim Dataset V1, one of the largest open‑source simulation datasets for Franka arm manipulation. The dataset, training scripts, and benchmark results are all available for free download.

Dataset scale and composition

V1 contains more than 50,000 human‑teleoperated simulation trajectories covering 207 manipulation tasks such as pick‑and‑place, stacking, pouring, articulated‑object handling, and tool use. These trajectories span over 60,000 scene variants generated on a simulated Franka Research 3 arm. The data were collected via Axis’s browser‑based teleoperation platform, Axis Hub, by a distributed crowd of contributors.

Benchmark performance

Continual pre‑training on V1 improves the LIBERO‑Plus benchmark metric π0.5 from 83.9 % to 88.8 % success, surpassing a volume‑matched RoboCasa365 baseline by 37.3 %. Performance gains increase consistently as the proportion of the dataset used for pre‑training grows from 25 % to 100 %, with the largest improvements observed under camera, sensor‑noise, and layout perturbations.

Compounding data engine

Axis describes the dataset as part of a larger “compounding data engine” that iteratively determines new data collection needs based on model performance and failure cases. The engine operates across four data lines:

• Simulation: Over 200,000 contributors on Axis Hub have generated more than 4.7 million trajectories across 13 robot embodiments.
• Egocentric capture: A network of 1,000+ full‑time collectors records first‑person activity in real homes and businesses, accumulating over 200,000 hours.
• Loco‑manipulation: 500+ hours of combined mobility and dexterity data from humanoid platforms such as Unitree G1 and Booster T2.
• Human‑gated DAgger: 500+ hours of human‑in‑the‑loop correction targeting edge‑case deployments.

All data are recorded on‑chain on Base for provenance, and contributors receive rewards based on verified quality.

Commercial collaborations

Beyond open‑source release, Axis partners with robot embodiment companies to create customized data pipelines and model priors. For example, with Booster Robotics, Axis generated 42,000 simulation episodes on a digital twin of Booster’s workspace, producing a model prior that achieved 87.5 % success with only 30 real‑robot demonstrations, compared to 37.5 % for an out‑of‑the‑box baseline.

Funding and future plans

Axis Robotics raised $12 million in seed funding led by Hack VC, with participation from Nomad Capital, Pi Network Ventures, 10K Ventures, and angel investors. The company is already developing Dataset V2, targeting 1.2 million trajectories across 1,200 tasks and cross‑embodiment generalization.