Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Dual Process VLA: Efficient Robotic Manipulation Leveraging VLM

ByungOk Han, Jaehong Kim, Jinhyeok Jang

TL;DR

The paper presents DP-VLA, a hierarchical Vision-Language-Action framework that splits reasoning and control into a slow, high-level Large System 2 (L-Sys2) and a fast Small System 1 (S-Sys1). By extracting latent features with a VLM/VLA-based L-Sys2 at low frequency and using a lightweight S-Sys1 policy for real-time actions, the approach reduces computational burden while maintaining or improving task success. Experiments on RoboCasa show DP-VLA outperforms OpenVLA-fine-tuned and BC-Transformer baselines in both task success and inference speed, aided by ablations highlighting the value of decoding-stage latent features and the benefits of using pretrained L-Sys2 features. The work highlights a practical path toward efficient, scalable VLA systems for real-time robotic manipulation.

Abstract

Vision-Language-Action (VLA) models are receiving increasing attention for their ability to enable robots to perform complex tasks by integrating visual context with linguistic commands. However, achieving efficient real-time performance remains challenging due to the high computational demands of existing models. To overcome this, we propose Dual Process VLA (DP-VLA), a hierarchical framework inspired by dual-process theory. DP-VLA utilizes a Large System 2 Model (L-Sys2) for complex reasoning and decision-making, while a Small System 1 Model (S-Sys1) handles real-time motor control and sensory processing. By leveraging Vision-Language Models (VLMs), the L-Sys2 operates at low frequencies, reducing computational overhead, while the S-Sys1 ensures fast and accurate task execution. Experimental results on the RoboCasa dataset demonstrate that DP-VLA achieves faster inference and higher task success rates, providing a scalable solution for advanced robotic applications.

A Dual Process VLA: Efficient Robotic Manipulation Leveraging VLM

TL;DR

The paper presents DP-VLA, a hierarchical Vision-Language-Action framework that splits reasoning and control into a slow, high-level Large System 2 (L-Sys2) and a fast Small System 1 (S-Sys1). By extracting latent features with a VLM/VLA-based L-Sys2 at low frequency and using a lightweight S-Sys1 policy for real-time actions, the approach reduces computational burden while maintaining or improving task success. Experiments on RoboCasa show DP-VLA outperforms OpenVLA-fine-tuned and BC-Transformer baselines in both task success and inference speed, aided by ablations highlighting the value of decoding-stage latent features and the benefits of using pretrained L-Sys2 features. The work highlights a practical path toward efficient, scalable VLA systems for real-time robotic manipulation.

Abstract

Vision-Language-Action (VLA) models are receiving increasing attention for their ability to enable robots to perform complex tasks by integrating visual context with linguistic commands. However, achieving efficient real-time performance remains challenging due to the high computational demands of existing models. To overcome this, we propose Dual Process VLA (DP-VLA), a hierarchical framework inspired by dual-process theory. DP-VLA utilizes a Large System 2 Model (L-Sys2) for complex reasoning and decision-making, while a Small System 1 Model (S-Sys1) handles real-time motor control and sensory processing. By leveraging Vision-Language Models (VLMs), the L-Sys2 operates at low frequencies, reducing computational overhead, while the S-Sys1 ensures fast and accurate task execution. Experimental results on the RoboCasa dataset demonstrate that DP-VLA achieves faster inference and higher task success rates, providing a scalable solution for advanced robotic applications.

Paper Structure

This paper contains 15 sections, 3 equations, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Motivation
  3. Method: Dual Process VLA
  4. Large System 2 Model
  5. Small System 1 Model
  6. Experiments
  7. Experiment Details
  8. Comparisons with Other Methods
  9. Task Success Rate
  10. Inference Speed
  11. Ablation Study
  12. Comparisons across Latent Feature Types
  13. Analysis about the Effectiveness of Finetuning
  14. Future Work
  15. Conclusion

Figures (2)

  • Figure 1: Our Dual Process framework integrates the concepts of System 2 and System 1. The Large System 2 Model (L-Sys2) extracts latent features that encode both reasoning information related to user instructions and environmental context. These high-level representations guide the Small System 1 Model (S-Sys1) in generating fine-grained actions in real-time, leveraging various observations and the robot's states.
  • Figure 2: RoboCasa simulation environment. A Franka Emika Panda robot was used for training and evaluation across diverse kitchen settings.