I recently reviewed AxisRobotics’ progress from September to early October. It’s now less about building another robot model and more about creating a self-sustaining data loop. For Physical AI, models and compute are rarely the first bottlenecks. The real challenge lies in obtaining sufficient, diverse operational data—and a robust evaluation system that can consistently identify where the model falls short. Axis has integrated data collection into browser-based simulations, allowing users to remotely control robotic arms. Over 6 million trajectories have been collected, with more than 200,000 contributors; according to the team, their engine can generate around 100,000 new trajectories per day. The community sale saw 2,472 participants commit approximately 2.39 million USDC—roughly 2.39x oversubscribed—making it Sonar’s most participated sale this year. Numbers alone don’t prove the model can perform real tasks. What they’re aiming for is a different pipeline: strategies are executed in simulation, humans correct failures, and those corrections become new training trajectories. Evaluations aren’t locked into static benchmarks either—the Open Axis Benchmark dynamically samples new tasks from an existing library of over 6,000 tasks and 5.5 million trajectories. Their related paper has been accepted by the Physical World Models workshop at IROS 2026. They report that feeding back previously deployed rollouts improved real-robot success rates from 22% to 52%; specialized policies achieve near-perfect success under test conditions, with single-task compute costs estimated at $5–$10. These figures currently come primarily from the team; external reproducibility remains to be verified. Scale is still very early. According to their estimates, general capability will require around 100 million hours of training data; they currently have about 50,000 hours—less than one-thousandth of that target. While centralized labs scale linearly with headcount, open networks offer an alternative path—but it’s far from proven. My entry point is: https://t.co/eT4RE35jjM What matters most in October isn’t whether trajectory counts hit another zero—it’s whether this closed loop can elevate the next version of their strategy.