The Data Bottleneck in Robot Learning
Effective robot manipulation policies depend on diverse, high-quality demonstrations — yet scaling data collection has remained a persistent challenge. Existing pipelines often rely on specialized hardware, centralized operators, or fixed task suites, making it difficult to grow datasets organically as models improve.
The AXIS Platform
AXIS addresses this bottleneck with a community-driven data engine that enables browser-based teleoperation, allowing anyone with a web browser to contribute demonstrations. The platform automatically generates and validates new manipulation tasks, then transforms community-collected demonstrations into training-ready data through automated success checking, quality filtering, trajectory smoothing, and visual plus physics-based augmentation.
The current AXIS dataset contains 207 diverse tasks and more than 50,000 trajectories. Tasks span single-arm manipulation, dual-arm coordination, tool use, and long-horizon procedural tasks, organized into task snapshots with systematic held-out evaluation protocols.
Key Findings
The research team evaluated vision-language-action (VLA) policies under a unified AXIS evaluation suite and analyzed scaling behavior across different data volumes. The results demonstrate clear scaling effects:
- Continual pretraining on AXIS improves the overall success rate of π0.5 by 5.8% compared to the baseline
- AXIS-pretrained models outperform those pretrained on RoboCasa365 by 37.3%
- The largest gains appear under layout perturbations, sensor noise, and camera variations — conditions that most closely reflect real-world deployment challenges
- Success rates scale consistently with increasing data volume, suggesting continued headroom for community-driven data growth
Broader Implications
The AXIS approach represents a paradigm shift in how robot learning data is collected and curated. By lowering the barrier to contribution from specialized hardware to any web browser, the platform taps into a vastly larger pool of human demonstrators. This community-driven model could accelerate progress in embodied AI in much the same way that open-source datasets transformed computer vision and natural language processing a decade ago.
The project is open-source, with code, dataset, and benchmark available on the project website, inviting the broader robotics community to both contribute demonstrations and evaluate their policies on the AXIS benchmark.
