Robot Data Filtering Sped Up 313x

jiqizhixin · x · 2026-07-11

This work introduces ATHENA to make robot data annotation/filtering more efficient by using influence functions to determine which demonstration data actually aids task performance.
Methodologically, it leverages neural network gradient structures and uses rank-r random approximation to estimate the Hessian, making previously slow computations viable.
Results show a roughly 313x speedup in influence calculations; in simulations, using only 50% of the demonstration data matched or exceeded full fine-tuning performance.
Across 6 real-world robot tasks, achieving full-data performance required only 66.7% of the data.

More from Embodied