Axis Robotics Launches Open-Source Franka Arm Simulation Dataset for AI Development

Axis Robotics has announced the release of the Axis Sim Dataset V1, which is recognized as one of the largest open-source simulation datasets for Franka arm manipulation. This dataset includes the full dataset, training code, and benchmarks, all made publicly available.

The Axis Sim Dataset V1 comprises more than 50,000 human-teleoperated simulation trajectories that cover 207 manipulation tasks and over 60,000 scene variants, all based on the simulated Franka Research 3 arm. This dataset has achieved over 160,000 downloads, making it the most downloaded open-source simulation dataset for Franka manipulation on Hugging Face.

In benchmark tests, continual pretraining on the dataset has shown significant improvements, lifting the π0.5 metric and surpassing a volume-matched RoboCasa baseline. Axis Robotics aims to create a comprehensive data engine for Physical AI, integrating large-scale simulation with real-world data capture and humanoid loco-manipulation.

Axis Robotics has raised $12 million in seed funding, led by Hack VC, with contributions from various investors including Nomad Capital and Pi Network Ventures. The company challenges the traditional notion that only optimal data is useful for training, arguing that diverse and noisy data can lead to effective learning outcomes.

The dataset was developed in collaboration with researchers from esteemed institutions such as UC Berkeley, Johns Hopkins, and the University of Michigan. Future developments include scaling the dataset to 1.2 million trajectories across 1,200 tasks, further enhancing the robustness of the training policies.

Axis Robotics is also working with other companies to create customized data pipelines and model priors, demonstrating the practical applications of their open-source dataset in commercial robotics.