FOCA: Future-Oriented Conditioning for Data-Efficient Vision-Language-Action Adaptation

FOCA (Future-Oriented Conditioning) is a new framework for enhancing data-efficient adaptation in Vision-Language-Action (VLA) models, addressing their limitations in few-shot imitation learning. It integrates future interaction embeddings with goal-oriented observations, facilitating long-horizon reasoning without requiring pixel-level predictions. Experimental results indicate FOCA achieves a 95.7% success rate with only 20 demonstrations on the LIBERO benchmark, improves performance by 7-12% on RoboCasa, and provides up to 26% absolute gains on real robots, setting a new benchmark in few-shot VLA adaptation, which is crucial for practitioners aiming to optimize robotic control with limited data.