ActDistill: General Action-Guided Self-Derived Distillation for Efficient Vision-Language-Action Models

2025-11-22

AI Summary

Current Vision-Language-Action (VLA) models suffer from high computational overhead and excessive inference latency, hindering real-time deployment in robotic manipulation. To address this, we propose an action-guided self-distillation framework—the first efficient VLA model compression paradigm explicitly optimized for action execution. Our method introduces an action-prior-driven knowledge distillation mechanism, coupled with a graph-structured encapsulation module and a dynamic routing architecture, enabling precise transfer of the teacher’s action-decision capability to a lightweight student model. It integrates hierarchical supervision, dynamic computation path selection, and structured knowledge representation. Evaluated on embodied AI benchmarks, the distilled student model matches or surpasses the full-scale teacher in task performance while reducing computational cost by over 50% and accelerating inference by 1.67×.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Learning on the Edge & Model Compression
• Intelligent Robots: Manipulation

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Recent Vision-Language-Action (VLA) models have shown impressive flexibility and generalization, yet their deployment in robotic manipulation remains limited by heavy computational overhead and inference latency. In this work, we present ActDistill, a general action-guided self-derived distillation framework that transfers the action prediction capability of any existing VLA model to a lightweight counterpart. Unlike previous efficiency strategies that primarily emphasize vision-language correlations, ActDistill leverages action priors to guide knowledge transfer and model compression, achieving action-oriented efficiency for VLA models. Specifically, we employ a well-trained VLA model as the teacher and introduce a graph-structured encapsulation strategy to explicitly model the hierarchical evolution of action prediction. The student model, derived from the graph-encapsulated teacher, is further equipped with a dynamic router that adaptively selects computation paths based on action prediction demands, guided by hierarchical graph-informed supervision to ensure smooth and efficient evolution. During inference, graph-related auxiliary components are removed, allowing the student to execute only dynamically routed layers and predict high-precision actions with minimal computation and latency. Experiments on embodied benchmarks demonstrate that ActDistill achieves comparable or superior performance to full-scale VLA models while reducing computation by over 50% with up to 1.67 times speedup, thereby establishing a general paradigm toward efficient embodied intelligence.

Problem

• Reduces computational overhead in vision-language-action models for robotics
• Transfers action prediction capabilities to lightweight models via distillation
• Enables efficient inference with dynamic routing and hierarchical supervision

Innovation

• Action-guided distillation transfers VLA capabilities to lightweight models
• Graph-structured encapsulation models hierarchical action prediction evolution
• Dynamic router selects computation paths based on action demands