Axis robotics launches browser-based data engine for robot manipulation

Axis Robotics, in collaboration with UC Berkeley, Georgia Tech, and NTU, has introduced AXIS, a browser-based data engine designed to address the limitations of traditional robot manipulation datasets. Unlike conventional methods that rely on centralized data collection, AXIS enables real-time demonstration collection through a web interface, with processing handled by backend GPUs. The system supports 207 tasks and 50,129 trajectories, with data hosted on Hugging Face and accessible for non-commercial academic use. AXIS leverages a MuJoCo WebAssembly frontend for teleoperation, allowing users to control a Franka Research 3 robot with a parallel-jaw gripper using various input methods. The dataset includes detailed metadata, such as task success labels and RGB-D observations. Cleaning processes, including motion smoothing and resampling, enhance data quality, though they result in a slight drop in replay success rates. The system also supports pretraining and fine-tuning on LIBERO, achieving improved performance metrics compared to baseline models. The project offers a comprehensive platform for researchers and developers, with access to the paper, dataset, and project page. It also invites community engagement through social media and newsletters, emphasizing its open-source and collaborative approach to advancing robot manipulation research.