Visual-language-action model optimization method based on actionable neighborhood prior

By using a vision-language-action model optimization method based on action neighborhood priors, the problems of overfitting and low sample efficiency of existing VLA models in robot operation are solved, achieving higher robustness and generalization ability, and applicable to scenarios such as industrial flexible manufacturing, logistics warehousing and sorting, home service robots and special environment operations.

CN122153840APending Publication Date: 2026-06-05SHANGHAI JIAOTONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
SHANGHAI JIAOTONG UNIV
Filing Date
2026-03-09
Publication Date
2026-06-05

AI Technical Summary

Technical Problem

Existing vision-language-action (VLA) models ignore the feasible neighborhood and approximate equivalence features of physical actions in robot operations, resulting in overfitting in the supervised fine-tuning stage, low sample efficiency in the reinforcement fine-tuning stage, and an inability to adapt to the accuracy requirements under different states, as well as insufficient generalization ability.

Method used

An optimization method based on Feasible Action Neighborhood Prior (FAN) is adopted. The action space is discretized by multidimensional binning, and a Gaussian distribution is introduced to approximate the neighborhood geometry. In the supervision and reinforcement fine-tuning, KL divergence regularization term and FAN-PPO loss function are introduced to dynamically adjust the action distribution to match the task requirements.

Benefits of technology

It significantly improves the robot's robustness and generalization ability in complex environments, reduces dependence on teaching data, improves training stability and sample efficiency, reduces engineering costs, and is suitable for a variety of general robot operation scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122153840A_ABST
    Figure CN122153840A_ABST
Patent Text Reader

Abstract

The application discloses a visual-language-action model optimization method based on an actionable neighborhood prior, and the method comprises the following steps: collecting operation trajectory data containing natural language instructions, visual observation and continuous control action sequences and preprocessing; adopting a self-recurrence structure of an image encoder, a language encoder and an action decoder, discretizing a continuous action space through multidimensional binning; defining an actionable neighborhood based on a value function constraint, modeling the neighborhood by using a Gaussian distribution, introducing a KL divergence regularization term and a negative log-likelihood loss to construct a supervised fine-tuning framework; adaptively adjusting the width and narrowness of the action distribution by using an AdaFAN algorithm, and introducing geometric constraints into a FAN-PPO algorithm in a proximal policy optimization framework to reduce invalid exploration; and deploying the model to realize closed-loop control. The application is compatible with mainstream VLA models and hardware, significantly enhances the adaptability of the model to cumulative visual observation disturbance and the training stability, and is suitable for various robot operation scenes such as industrial flexible manufacturing and home service.
Need to check novelty before this filing date? Find Prior Art