ATHENA: Accelerated Multi-Task Heterogeneous Influence Functions for Robot Data Curation

2026-06-15

AI Summary

This work addresses the challenge of scaling influence functions to billion-parameter vision-language-action (VLA) models in robotic imitation learning, where existing methods suffer from prohibitive computational costs and multi-task coordination bottlenecks. The authors propose ATHENA, the first framework to enable efficient influence estimation for large-scale multi-task VLA models. ATHENA reduces projection overhead by decomposing linear-layer gradients using Kronecker structure and accelerates Hessian inverse computation via rank-r randomized truncation. By integrating global and local interaction-aware influence mechanisms, it effectively balances data selection across 50 joint tasks. Experiments on RoboTwin 2.0 and real robots demonstrate that ATHENA achieves superior performance using only 50% of simulated or 66.7% of real demonstration data compared to full fine-tuning, while accelerating influence computation by 313.4×.

Abstract

In robot imitation learning, influence functions provide a principled approach to quantify each demonstration's effect on robot task outcomes, yet scaling them to billion-parameter Vision-Language-Action (VLA) models is limited by computational and multitask bottlenecks. To this end, we propose ATHENA, an influence function framework tailored for multitask VLA data curation at a billion-parameter scale. Concretely, it leverages the Kronecker structure of linear-layer gradients to reduce projection cost, and approximates dense Hessian inversion with a rank-r Random Truncated Approximation, achieving about a 313.4x speedup in influence computation. Furthermore, ATHENA formulates global and local interactive influence to balance data curation across 50 jointly trained tasks. Extensive evaluations on RoboTwin 2.0 and real-robot deployment, covering 9.34 and 6.90 hours of demonstrations, respectively, show that ATHENA matches or exceeds full-data joint fine-tuning using only 50% of demonstrations in simulation and 66.7% of data across six real-robot tasks. Overall, ATHENA demonstrates its effectiveness for data curation in billion-parameter multitask VLA fine-tuning.

Problem

• influence functions
• robot imitation learning
• Vision-Language-Action models
• data curation
• multitask learning

Innovation

• Influence Functions
• Vision-Language-Action Models
• Multi-Task Learning
• Data Curation
• Hessian Approximation