Franka arm simulation dataset tops 160,000 downloads, becomes most downloaded

Axis Robotics has thrown open the doors on one of the largest open-source collections of robot manipulation data ever released, betting that scale and diversity — not curated perfection — is what will actually teach machines to move through the physical world. The company’s new Franka arm simulation dataset, called Axis Sim Dataset V1, arrives with full training code and benchmarks publicly available, and it’s already reshaping how researchers think about training data for what the industry calls Physical AI.

Key takeaways

• Axis Sim Dataset V1 contains over 50,000 human-teleoperated trajectories across 207 manipulation tasks and 60,000+ scene variants on a simulated Franka Research 3 arm.
• The dataset has logged more than 160,000 downloads on Hugging Face, the most of any open-source Franka simulation dataset.
• Pretraining on V1 boosts π0.5 performance on LIBERO-Plus and beats a volume-matched RoboCasa baseline by 37.3%.
• Axis Robotics raised $12 million in seed funding led by Hack VC, with Nomad Capital, Pi Network Ventures, and 10K Ventures also participating.
• A future V2 release is planned to scale to 1.2 million trajectories across 1,200 tasks.

Axis Robotics Launches Major Franka Arm Simulation Dataset

Axis Robotics built V1 as a direct challenge to a long-standing assumption in robotics research: that training data has to be clean, expert-generated, and tightly controlled to be useful. Instead, the company released a dataset assembled from more than 50,000 human-teleoperated trajectories, spanning 207 distinct manipulation tasks and over 60,000 scene variants, all captured on a simulated Franka Research 3 arm.

Axis Sim Dataset V1 scale and access

The full dataset, along with training code and benchmark results, is publicly available for anyone to download and test. That openness is part of the point — Axis wants outside researchers to verify its results rather than take the company’s word for it.

Dataset popularity and benchmarking results

The uptake has been notable. V1 has drawn more than 160,000 downloads on Hugging Face, making it the most downloaded open-source simulation dataset built around Franka arm manipulation. On the benchmark side, continual pretraining on V1 lifted the π0.5 model from 83.9% to 88.8% success on LIBERO-Plus and outperformed a volume-matched RoboCasa baseline by 37.3%. Performance kept climbing as more of the dataset was used in pretraining, from 25% up to 100%, without any sign of leveling off — a signal, Axis argues, that the gains come from genuine diversity in the data rather than a one-time boost.

Innovative Data Collection and Processing Approach

Most robotics teams collect data by filtering down to a small pool of expert demonstrations and discarding anything that looks noisy. Axis does the opposite, and that choice is the real story behind this open source robotics dataset.

Distributed crowd teleoperation over filtered expert data

The company’s argument is that data quality lives at the level of the whole distribution, not any single trajectory. When a large, diverse crowd generates messy, imperfect demonstrations and their mistakes aren’t correlated with each other, the noise cancels out during training and a workable policy still emerges. V1 puts that theory into practice: its trajectories cover pick-and-place, stacking, pouring, articulated-object handling, and tool use, all gathered through Axis’s browser-based teleoperation platform, Axis Hub, using a distributed crowd rather than one in-house expert team. Researchers from UC Berkeley, Johns Hopkins, and the University of Michigan contributed to building the dataset.

Axis’s Physical AI data engine components

V1 is just one output of a larger machine. Axis describes its approach as a data engine that keeps producing new training material rather than a static download that goes stale. It runs on four fronts at once: large-scale simulation, egocentric real-world capture, humanoid loco-manipulation, and human-gated DAgger post-training, where people step in to correct edge cases the model struggles with. Model failures feed back into what gets collected next, so each training cycle informs the following one.

Partnerships and On-Chain Data Provenance

Beyond the open dataset, Axis is already turning this pipeline into commercial deployments — and that’s where the practical stakes of this work become clearer.

Collaboration with Booster Robotics and other partners

As Booster Robotics’ first simulation-data partner, Axis rebuilt Booster’s actual workspace as a task-specific digital twin, then had distributed contributors collect over 42,000 simulation episodes on it. Those episodes were distilled into a model prior built specifically for Booster’s hardware. The payoff: with just 30 real-robot demonstrations, that customized prior hit 87.5% success, compared to 37.5% for an out-of-the-box π0.5 model — matching π0.5’s performance while using half the real-world demonstration data. Axis also works with other embodiment and model companies, alongside industrial automation players, and supplies data to on-chain robotics networks.

On-chain data provenance and contributor rewards

Every task and trajectory collected through Axis Hub gets recorded on-chain on the Base blockchain, creating a verifiable record of where the data came from. Contributors are rewarded for work that passes quality checks, an incentive structure meant to keep the crowd-sourced model honest at scale. This is where the broader relevance of the Axis Robotics Physical AI approach starts to show: provenance tracking isn’t just a technical footnote, it’s the mechanism that lets a distributed, noisy crowd substitute for a small team of trusted experts without losing accountability.

Future Dataset Expansion and Vision for Physical AI

Axis isn’t treating V1 as a finished product — it’s a checkpoint. The company says a V2 release is already underway, aiming to scale the dataset to 1.2 million trajectories across 1,200 tasks, with a focus on cross-embodiment generalization so models trained on the data transfer better across different robot bodies.

Founder’s vision on continuous data production

Chris Feng, founder of Axis Robotics, framed the company’s long-term bet this way: “The future of Physical AI isn’t a static dataset you download once. 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, alongside serial founders who have previously scaled consumer platforms to more than 30 million users. Jiachen Li, an Assistant Professor at Georgia Tech, advises the company’s research.

Funding and Research Advisement

Axis Robotics raised $12 million in seed funding to build out this data engine, with Hack VC leading the round. Nomad Capital, Pi Network Ventures, 10K Ventures, and a group of angel investors also participated. The backing gives the company runway to keep expanding its four data lines — simulation, egocentric capture, loco-manipulation, and human-gated correction — while pushing toward the larger V2 release.

Why this matters for the wider field: robotics researchers have spent years assuming that training data needs to be near-perfect to be usable. If Axis’s results hold up outside its own benchmarks, it suggests robust policies can be trained on noisy, crowd-sourced data at a fraction of the cost of hiring expert demonstrators — which could reshape how every lab and startup building humanoid or arm-based robots sources its training data going forward.

FAQ

What is the Axis Sim Dataset V1?

Axis Sim Dataset V1 is one of the largest open-source simulation datasets for Franka arm manipulation, with over 50,000 human-teleoperated trajectories spanning 207 tasks.

How does Axis Robotics gather data for their simulation datasets?

Axis Robotics collects data from a distributed crowd using its browser-based teleoperation platform, Axis Hub, rather than relying solely on filtered expert demonstrations.

What benefits does pretraining on the Axis Sim Dataset V1 provide?

Pretraining on V1 improves model success rates on manipulation benchmarks such as LIBERO-Plus, outperforming baseline datasets like RoboCasa by a wide margin.

What is Axis Robotics’ vision for the future of Physical AI datasets?

Axis Robotics envisions Physical AI as a continuous data engine that keeps producing new training data to meet model needs, emphasizing scale, diversity, and closed-loop improvement rather than a one-time dataset release.

Article produced with the assistance of artificial intelligence and reviewed by the editorial team.