Cartwheel reveals compute-optimal motion scaling

This moves human-motion generation from “scale and hope” toward predictable engineering. The result is highly relevant to animation and embodied AI, though independent replication remains essential.

• –Both autoregressive and flow-matching models reach near-C^1/2 compute-optimal frontiers, suggesting the behavior is not tied to one architecture.
• –The final corpus contains 11,969 hours and 5.56 million clips, making data construction a central research contribution.
• –At matched model size and compute, flow matching reportedly tolerates roughly 30× more data reuse than autoregression before held-out loss degrades.
• –Developers can use these curves to plan model size, dataset growth, and training budgets before committing major GPU resources.
• –Scaling loss predictability does not automatically guarantee physical realism, controllability, or successful transfer to robots.