Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize Better

TL;DR

This paper introduces a method to preserve the semantic knowledge of large vision-language models during training of vision-language-action systems, enabling faster training, real-time inference, and better generalization for robotic control.

Contribution

It proposes a knowledge insulation technique that maintains pretrained VLM knowledge during VLA training, improving training speed and control performance.

Findings

Naive inclusion of action experts harms training and knowledge transfer.

Knowledge insulation mitigates knowledge degradation during training.

The method enhances real-time control and generalization in robotic systems.

Abstract

Vision-language-action (VLA) models provide a powerful approach to training control policies for physical systems, such as robots, by combining end-to-end learning with transfer of semantic knowledge from web-scale vision-language model (VLM) training. However, the constraints of real-time control are often at odds with the design of VLMs: the most powerful VLMs have tens or hundreds of billions of parameters, presenting an obstacle to real-time inference, and operate on discrete tokens rather than the continuous-valued outputs that are required for controlling robots. To address this challenge, recent VLA models have used specialized modules for efficient continuous control, such as action experts or continuous output heads, which typically require adding new untrained parameters to the pretrained VLM backbone. While these modules improve real-time and control capabilities, it remains an open question whether they preserve or degrade the semantic knowledge contained in the pretrained VLM, and what effect they have on the VLA training dynamics. In this paper, we study this question in the context of VLAs that include a continuous diffusion or flow matching action expert, showing that naively including such experts significantly harms both training speed and knowledge transfer. We provide an extensive analysis of various design choices, their impact on performance and knowledge transfer, and propose a technique for insulating the VLM backbone during VLA training that mitigates this issue. Videos are available at https://pi.website/research/knowledge_insulation.