arXiv:2511.18082v1 Announce Type: new 
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.

تم تقديم ActDistill كإطار عام للتقطير الذاتي الموجه بالإجراء، يهدف إلى تحسين كفاءة نماذج اللغة-الرؤية-الإجراء (VLA). تركز هذه المقاربة المبتكرة على نقل قدرات التنبؤ بالإجراء من نموذج VLA مدرب جيدًا إلى نسخة خفيفة الوزن، مما يعالج الحمل الحسابي ووقت الاستدلال الذي يحد من تطبيقات المناورة الروبوتية.

ActDistill se ha presentado como un marco general de destilación auto-dirigida guiada por la acción, con el objetivo de mejorar la eficiencia de los modelos de Acción-Lenguaje-Visión (VLA). Este enfoque innovador se centra en transferir las capacidades de predicción de acción de un modelo VLA bien entrenado a una versión más ligera, abordando la sobrecarga computacional y la latencia de inferencia que limitan las aplicaciones de manipulación robótica.

ActDistill a été introduit comme un cadre général de distillation auto-dérivée guidée par l'action, visant à améliorer l'efficacité des modèles Vision-Language-Action (VLA). Cette approche innovante se concentre sur le transfert des capacités de prédiction d'action d'un modèle VLA bien entraîné vers une version allégée, abordant ainsi la surcharge computationnelle et la latence d'inférence qui limitent les applications de manipulation robotique.

ActDistill has been introduced as a general action-guided self-derived distillation framework aimed at enhancing the efficiency of Vision-Language-Action (VLA) models. This innovative approach focuses on transferring action prediction capabilities from a well-trained VLA model to a lightweight version, addressing the computational overhead and inference latency that limit robotic manipulation applications.

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

One More Thing in AI – Your Shortcut to AI Mastery

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

Was this article worth reading? Share it

One More Thing in AI

LucidQuery AI

The Visualizer

Dyad

Legion AI

LangWatch

Ready to build your own newsroom?