A new scaling paradigm for robotic model training: from raw trajectories to reusable, composable and distillable skills.
Read the full blog:
Three findings from our latest experiments:
• Embodied RSI: the deployed policy generates real-world data that trains its stronger successor (22% to 52%).
• A lightweight proxy model scores trajectories by utility and coverage, so a filtered subset can beat the full pool.
• Expert models reach near 100% success for $5 to $10 in compute per task and can be chained into long-horizon behaviors.
What Axis scales is a continuously expanding library of robotic capabilities that can be produced, composed and distilled into general-purpose models.
显示更多