AXIS technical analysis @axisrobotics Axis is building a closed-loop data engine for Physical AI, not just another robotics dataset Contributors teleoperate a Franka arm in-browser via MuJoCo-WASM. Every accepted trajectory is scored, cleaned, domain-randomized in Isaac Sim, and recorded on Base with a unique Data ID Once a policy is trained, the same task reopens as a post-training session. The policy runs until it fails; the contributor takes over (human-gated DAgger), and those corrections become the highest-value training signal. The improved policy then decides which new scenes and failure modes to collect next This creates a compounding loop: model weakness → targeted human correction → stronger policy → new data requirements Early results show the effect. Continual pretraining on Axis data lifted π0.5 from 83.9 to 88.8 on LIBERO-Plus. In a real deployment, feeding failure data back raised success from 22% to 52% The bottleneck was never model architecture. It was scalable, high-quality, on-policy data. Axis is turning that bottleneck into a live, global data flywheel.