The dataset was collected using a simulated Franka Research 3 arm and covers more than 60,000 scene variations. Tasks include pick-and-place, stacking, pouring, tool use and manipulating articulated objects. The full dataset, training code and benchmarks have been made publicly available.
Rather than relying only on expert demonstrations, Axis collected the data through its browser-based teleoperation platform using a distributed group of human operators. The company is testing whether larger amounts of varied, imperfect demonstrations can provide useful training data for general-purpose robot models.
Axis tested the dataset by continually pretraining Physical Intelligence’s π0.5 model. On the LIBERO-Plus benchmark, success increased from 83.9% to 88.8%. Axis also reports that the model outperformed a volume-matched RoboCasa baseline, with performance continuing to improve as more of the dataset was added.
The dataset has already recorded more than 160,000 downloads on Hugging Face. Axis says a second version is now being developed with a target of 1.2 million trajectories across 1,200 tasks.
Axis is also working beyond simulated robot arms. Its broader data platform covers real-world egocentric data collection, humanoid loco-manipulation and human-guided post-training as it builds datasets for different robot embodiments.
The company raised $12 million in seed funding led by Hack VC, with participation from Nomad Capital, Pi Network Ventures and 10K Ventures.
