Axis Robotics hits 1M trajectories

Crowdsourcing simulation-based robot data bypasses the slow and expensive hardware collection bottleneck, potentially accelerating general-purpose physical AI models.

• –Reaching one million trajectories validates a Web3-incentivized global network approach to data gathering
• –Replaying MuJoCo physics data in Isaac Sim allows for visual domain randomization and robust sim-to-real policies
• –Models trained on this crowdsourced dataset reportedly outperform standard benchmarks like RoboCasa on pi-0.5
• –Validating the pipeline on OpenVLA and ACT shows simulation can transfer to real-world Franka arms