Axis Robotics Launches AXIS Browser Data Engine

Axis Robotics, in collaboration with researchers from UC Berkeley, Georgia Tech, and NTU, has introduced AXIS, a novel browser-based data engine designed to accelerate the collection of robot manipulation datasets. This system addresses the historical bottleneck in AI model training, where dataset growth has lagged behind model development due to centralized and static data collection methods. AXIS shifts demonstration collection to a web interface, offloads computationally intensive processes to backend GPUs, and treats the dataset as a continuously expanding resource rather than a fixed benchmark. The system's training code is publicly available as a patch layer over OpenPI, and its teleoperation platform is accessible through any web browser. The current dataset, hosted on Hugging Face, comprises 2.36 terabytes of data and is restricted to non-commercial academic use. No policy checkpoints have been released.

The core architecture of AXIS is built on an asymmetric design. Contributors use a web-based frontend, powered by MuJoCo WebAssembly, to teleoperate a Franka Research 3 robot equipped with a parallel-jaw gripper. Input methods include keyboard, mouse, virtual joystick, or gamepad. Physics simulation and Three.js rendering occur on the React UI thread, ensuring that logged state-action samples remain synchronized with the simulator. All heavy processing, including rendering on eight RTX 4090 GPUs and training/evaluation on eight A100 GPUs, is handled by backend servers. The tasks themselves are procedurally generated by TaskGen, which decomposes language instructions into configurations for tasks, scenes, and objects. It retrieves or generates 3D meshes via an image-to-3D pipeline, scales them to realistic physical dimensions, and then proposes a 2.5D layout. A layout supervisor validates the generated scenes, repositioning, reorienting, or regenerating objects as needed to meet defined constraints. Each task is accompanied by a structured success checker that is re-executed on the backend, overriding any success flags from the frontend.

The initial release of the AXIS dataset contains 207 distinct robot manipulation tasks, encompassing 50,129 episodes and over 60,000 task or scene configurations. This extensive collection aims to provide a rich and dynamic resource for training AI models in complex robotic interactions. The system's approach democratizes data collection by enabling contributions from a wider community, moving away from the traditional model of expert operators in controlled lab environments. By making the teleoperation platform browser-based, Axis Robotics aims to foster a more collaborative and iterative development cycle for robot learning research. The decision to gate the dataset and restrict its use to academic purposes suggests a focus on foundational research and community building before broader commercial applications are considered. The ongoing expansion of the dataset is a key feature, promising to keep pace with the rapid advancements in AI model capabilities.