Axis Robotics has released Axis Sim Dataset V1, one of the largest open-source simulation datasets for Franka arm manipulation. The full dataset, training code, and benchmarks are now publicly available. V1 is built from over 50,000 human-teleoperated simulation trajectories across 207 manipulation tasks and more than 60,000 scene variations, all on a simulated Franka Research 3 arm. The dataset has attracted over 160,000 downloads, making it the most downloaded open-source simulation Franka manipulation dataset on Hugging Face. In benchmarks, continual pretraining on V1 improved ฯ0.5 performance and beat a volume-matched RoboCasa baseline, with all results open and verifiable.
Axis Robotics is building a compounding data engine for Physical AIโa fully integrated system that combines large-scale simulation, egocentric real-world capture, humanoid loco-manipulation, and human-gated DAgger post-training. The company has raised $12 million in seed funding, led by Hack VC, with participation from Nomad Capital, Pi Network Ventures, 10K Ventures, and angel investors.
A Bet Against “Clean Data Only”
A common assumption in robotics is that demonstrations must be near-optimal from the startโfilter down to expert trajectories, standardize the setup, and remove any noisy data before imitation learning can work. Axis takes the opposite view: data quality is determined at the distribution level, not by individual trajectories. When a large and diverse crowd produces noisy, suboptimal trajectories and their errors are uncorrelated, the noise cancels out, and a working policy emerges during training. Axis Sim Dataset V1 puts this idea to a public test. Its trajectories cover pick-and-place, stacking, pouring, articulated-object manipulation, and tool use, all collected through Axis’s browser-based teleoperation platform, Axis Hub, by a distributed crowd rather than a single expert team. The dataset was developed with researchers from UC Berkeley, Johns Hopkins, the University of Michigan, and other institutions.
Results That Scale
On LIBERO-Plus, continual pretraining on V1 lifts ฯ0.5 success from 83.9% to 88.8%, outperforming a volume-matched RoboCasa365 baseline by 37.3%. Performance improves steadily as pretraining data scales from 25% to 100% of the dataset, with no signs of saturationโevidence that the gains come from diversity and coverage, not a one-time boost. The largest improvements appear under camera, sensor-noise, and layout perturbations, which are exactly the variations Axis randomizes during generation. The team says 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 showing that suboptimal simulation data can train robust policies.
The Engine Behind the Dataset
The dataset is one output of a larger, continuously compounding data engine. While a traditional data vendor collects to a fixed specification and stops, Axis uses model performance and failure cases to decide what to collect next, so each training round informs the next. This engine runs on a hybrid strategy across four data lines, all now operating at scale:
– Simulation: Over 200,000 distributed contributors on Axis Hub, a top-3 dApp on Base, have produced 4.7 million+ trajectories across 13 embodiments.
– Egocentric: A managed network of 1,000+ full-time, QC-trained collectors captures first-person activity in real homes and businesses across 14 industries. Over 200,000 hours are already banked, growing by 4,000+ hours daily, with Vicon-verified hand pose.
– Loco-manipulation: 500+ hours combining mobility and dexterity on real humanoids (Unitree G1, Booster T2) through hardware-agnostic teleoperation.
– Human-gated DAgger post-training: 500+ hours of human-in-the-loop correction focused on deployment edge cases.
Every task and trajectory is recorded on-chain on Base for provenance, and contributors are rewarded for verified work quality.
From Open Data to Commercial Deployment
Beyond open-sourcing simulation data, Axis works directly with robot embodiment companies to build customized datasets for their specific needs.Axis builds custom data pipelines and model priors tailored to each robotics embodiment. As Booster Robotics’ first simulation-data partner, Axis recreated Booster’s real workspace as a task-focused digital twin. Distributed contributors then ran over 42,000 simulation episodes on it, which Axis distilled into a model prior specific to Booster’s needs. Using just 30 real-robot demonstrations, that prior achieved an 87.5% success rate, compared to 37.5% for a standard ฯ0.5 modelโmatching ฯ0.5’s performance with only half the real-world demos.
Axis’s other partners include embodiment companies like Feagine Robotics, model companies such as Manycore Tech and Dexmal, and industrial automation firms like Lotus Cars and Geely Auto. Axis also powers on-chain robotics networks, including 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 gives 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, along with serial founders who have scaled consumer platforms to over 30 million users. Its research is advised by Jiachen Li, Assistant Professor at Georgia Tech.
Paper Link: https://arxiv.org/abs/2607.21588
Project Page: https://axisaiorg.github.io/AXIS-V1/
Dataset Link: https://huggingface.co/datasets/axisrobotics/Franka-Dataset
Github Codebase: https://github.com/AxisAIOrg/Axis-V1-Training
Frequently Asked Questions
Here is a list of FAQs about Axis Robotics opensource Franka arm simulation dataset designed to be clear practical and useful for both newcomers and experienced roboticists
General Beginner Questions
Q What exactly did Axis Robotics release
A They released a massive opensource dataset of simulated data specifically for Franka robotic arms It contains millions of examples of the robot performing tasks in a simulated environment along with the sensor data that goes with it
Q Why is this a big deal for Physical AI
A Physical AI needs to understand how to interact with the real world Training these AI models usually requires a huge amount of realworld data which is slow and expensive to collect This dataset provides a massive highquality shortcut allowing AI models to practice billions of times in simulation first
Q What is a simulation dataset
A Its a collection of data generated inside a computer program rather than from a physical robot Think of it as a video game for robots where every action and observation is recorded to create training data for AI
Q Im new to robotics Can I use this dataset
A Absolutely Because its opensource you can download it and use it to learn You can start by exploring the data to understand what robotic sensor data looks like For handson use some basic Python and machine learning knowledge is helpful but the barrier to entry is much lower than needing your own physical robot
Q Is this dataset only for the specific Franka robot or can I use it for other robots
A The data is specific to the Franka arms movements and sensors However the techniques you learn and the AI models you train can often be adapted or transferred to other similar robot arms especially if you use a framework that supports model transfer
Technical Advanced Questions
Q What specific format is the dataset in
A The dataset is typically provided in a standard format like HDF5 or a similar structured format often aligned with common robot learning frameworks like RLBench or Robosuite It usually includes RGBD camera feeds depth maps robot joint states and action trajectories