Axis Robotics Launches AXIS Browser-Based Data Engine
Axis Robotics releases AXIS, a browser-based robot data engine featuring 207 tasks, 50,129 trajectories, and a 2.36 TB gated dataset.
Stock photo for illustration only, not from the actual event
- Axis Robotics released AXIS, a browser-based robot data engine.
- The Hugging Face dataset is 2.36 TB and restricted to academic use.
- It features 207 tasks and over 50,129 trajectories from 70,000+ contributors.
- The combined model achieves 88.8 overall on LIBERO-Plus benchmark.
Axis Robotics has introduced AXIS, a browser-based robot data engine. The training code is available as a patch layer over OpenPI, while the teleoperation platform runs live in any web browser. However, the dataset hosted on Hugging Face is gated at 2.36 TB, restricted strictly to non-commercial academic use, and no policy checkpoints have been released.
The core architectural decision behind the system is asymmetry. Contributors operate a Franka Research 3 robotic arm with a parallel-jaw gripper inside a MuJoCo WebAssembly frontend using a keyboard, mouse, virtual joystick, or gamepad. Physics stepping and Three.js rendering run independently from the React UI thread, ensuring logged state-action samples align with the simulator rather than the interface. Heavy computations are offloaded, with rendering handled by eight RTX 4090 GPUs and training and evaluation run on eight A100 GPUs.
Stock photo for illustration only, not from the actual event
Tasks are generated algorithmically rather than hand-authored. The TaskGen system decomposes a language instruction into task, scene, and object configurations, retrieves or generates meshes via an image-to-3D pipeline, rescales them to plausible physical sizes, and proposes a 2.5D layout. A layout supervisor validates the instantiated scene and repositions, reorients, or regenerates objects if constraints fail, while every task includes a structured success checker re-run by the backend.
The released snapshot contains 207 tasks, 50,129 episodes, and over 60,000 task or scene variants across seven scene categories. Each trajectory includes task metadata, embodiment details, simulator version, robot and object states, actions, success labels, and third-view plus wrist RGB-D observations, credited to more than 70,000 community contributors.
Releasing a browser-based data engine for robotics highlights the growing effort to lower infrastructure barriers and simplify data collection processes. By decoupling physics simulation and rendering from the user interface, researchers can more easily collaborate and scale up physical AI data acquisition through standard web browsers.
Data cleaning is treated as a rigorous production stage. Samples with joint variations below 5e-3 are dropped as static data, a Savitzky-Golay filter with a window of 15 and polynomial order 3 smooths continuous motion, and cubic splines resample web interface rates from 6-8 Hz up to a 20 Hz target.
All conditions initiate from the released π0.5 checkpoint, utilizing a PaliGemma Gemma-2B backbone paired with a Gemma-300M action expert. When combined with AXIS-100%, the model reaches an overall score of 88.8 on LIBERO-Plus, outperforming the vanilla π0.5 baseline score of 83.9.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment