Robot manipulation datasets have historically grown at a far slower pace than the models trained on them, largely due to closed, centralized collection pipelines. In conventional setups, expert operators gather demonstrations on lab hardware, process the data offline, and release a static benchmark that rarely expands. This creates a bottleneck: the datasets that drive progress in embodied AI remain limited and inflexible.
In response, a collaborative research team from Axis Robotics, UC Berkeley, Georgia Tech, NTU, and five other institutions has proposed a fundamental redesign of how such datasets are created and utilized. Their system, AXIS, shifts demonstration collection directly into the browser, offloads the heavy computational work to backend GPUs, and treats the dataset as a continuously growing resource rather than a one-time release.
This browser-based approach democratizes data collection—removing the need for specialized lab infrastructure and enabling broader participation. Meanwhile, backend GPUs handle the processing, ensuring scalability and efficiency. As a result, AXIS supports an expanding repository of robot manipulation tasks, currently encompassing 207 distinct tasks and 50,129 recorded trajectories, with the capacity to grow over time.
Deployment Status
AXIS is partially deployable as of September 2026. The training code is publicly available as a patch layer over OpenPI, a popular open-source policy learning framework. Additionally, the teleoperation platform is fully operational in any standard web browser, allowing users to contribute demonstrations from virtually anywhere. This dual approach—open-source training code and an accessible data collection interface—positions AXIS as a practical tool for both researchers and hobbyists alike.
Implications for 2026 and Beyond
As the robotics community moves toward more data-hungry models, particularly vision-language-action (VLA) systems, the need for dynamic, scalable datasets becomes critical. AXIS addresses this by offering a live, evolving dataset that can adapt to new tasks and environments without requiring a complete overhaul. This model may set a precedent for future data engines, shifting the focus from static benchmarks to continuous learning ecosystems.
For researchers, the architecture of AXIS—combining browser accessibility with backend compute—offers a template for scalable data collection in other domains beyond manipulation, such as navigation or human-robot interaction.
Availability
- Training Code: Publicly available as a patch over OpenPI
- Teleoperation Platform: Live at Axis Robotics Hub
- Dataset Access: Further details via the AXIS project page
via MarkTechPost