Figure May Have Found a Scaling Law for Robot Pretraining, Predicting 8x-Data Runs

rohanpaul_ai · x · 2026-09-18

Figure appears to have found a scaling rule for robot pretraining: with model and task training fixed, adding more of its Index human-behavior dataset makes action-prediction loss fall in a clean, predictable way. Smaller runs predicted the 8x-data run almost exactly, letting robotics firms estimate gains before spending compute — though the curve predicts action loss, not real-world task success, so predictable gains in robot reliability remain unproven.

Original post →

More from Embodied

Embodied channel →