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

This dataset drew over 160,000 downloads, making it the most downloaded open-source simulation Franka manipulation dataset on Hugging Face. In benchmarks, continual pretraining on V1 lifted π0.5 and beat a volume-matched RoboCasa baseline, with every result open and verifiable.

Axis Robotics is building the ultimate compounding data engine for Physical AI, a vertically integrated system spanning 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.

A Bet Against “Clean Data Only”

A common assumption in robotics is that demonstrations must be near-optimal to begin with — filter down to expert trajectories, standardize the setup, and discard anything noisy before it is safe to imitate. Axis’s thesis runs the other way: data quality lives at the distribution level, not the single trajectory. When a large and diverse enough crowd produces noisy, suboptimal trajectories and their errors are uncorrelated, the noise averages out and a working policy survives during training.

Axis Sim Dataset V1 puts that thesis to a public test. Its trajectories span 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 built 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 π0.5 from 83.9% to 88.8% success and outperforms a volume-matched RoboCasa365 baseline by 37.3%. Performance improves consistently as pretraining data scales from 25% to 100% of the dataset, with no saturation in sight, evidence that the gains come from diversity and coverage rather than a one-off bump. The largest improvements appear under camera, sensor-noise, and layout perturbations, the exact axes Axis randomizes during generation.

The team says V2 is already underway, 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 trains robust policies.

The Engine Behind the Dataset

The dataset is one output of a larger, actively compounding data engine. Where a traditional data vendor collects to a fixed spec and stops, Axis uses model performance and failure cases to determine what should be collected next, so every training round informs the next. That engine runs on a hybrid strategy across four data lines, and all four now run at scale:

• Simulation: over 200,000 distributed contributors on Axis Hub, a top-3 dApp on Base, producing 4.7M+ trajectories across 13 embodiments.
• Egocentric: a managed network of 1,000+ full-time, QC-trained collectors capturing first-person activity in real homes and businesses across 14 industries: 200,000+ hours already banked and growing by 4,000+ hours every day, 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 targeted at 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, embodiment-specific data pipelines and model priors.

As Booster Robotics’ first sim-data partner, Axis rebuilt Booster’s real workspace as a task-aligned digital twin, had distributed contributors collect 42,000+ simulation episodes on it, and distilled them into a Booster-specific model prior. With just 30 real-robot demos, that prior reached 87.5% success versus 37.5% for an out-of-the-box π0.5, matching π0.5 using half the real-world demonstrations.

Other partners span embodiment companies (Feagine Robotics), model companies (Manycore Tech, Dexmal) and industrial automation (Lotus Cars, Geely Auto). Axis also supplies on-chain robotics networks: 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, alongside 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:

Project Page:

Dataset Link:

Github Codebase:

Source: BeInCrypto