Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

From Knowing to Doing Precisely: A General Self-Correction and Termination Framework for VLA models

Wentao Zhang, Aolan Sun, Wentao Mo, Xiaoyang Qu, Yuxin Zheng, Jianzong Wang

TL;DR

VLA models struggle with precise manipulation and reliable task termination in open environments. The authors propose VLA-SCT, a lightweight, training-free self-correction framework that adds Trajectory Evaluation, Grasp Perturbation, and Termination Detection as an external control loop to existing VLA models. On LIBERO, VLA-SCT delivers state-of-the-art performance, achieving an average success rate of 81.55% with a 1.12× speedup over OpenVLA by improving both action precision and timely termination. This approach enhances robustness and deployment feasibility for VLA agents in complex, unstructured settings and suggests broad applicability across VLA architectures.

Abstract

While vision-language-action (VLA) models for embodied agents integrate perception, reasoning, and control, they remain constrained by two critical weaknesses: first, during grasping tasks, the action tokens generated by the language model often exhibit subtle spatial deviations from the target object, resulting in grasp failures; second, they lack the ability to reliably recognize task completion, which leads to redundant actions and frequent timeout errors. To address these challenges and enhance robustness, we propose a lightweight, training-free framework, VLA-SCT. This framework operates as a self-correcting control loop, combining data-driven action refinement with conditional logic for termination. Consequently, compared to baseline approaches, our method achieves consistent improvements across all datasets in the LIBERO benchmark, significantly increasing the success rate of fine manipulation tasks and ensuring accurate task completion, thereby promoting the deployment of more reliable VLA agents in complex, unstructured environments.

From Knowing to Doing Precisely: A General Self-Correction and Termination Framework for VLA models

TL;DR

VLA models struggle with precise manipulation and reliable task termination in open environments. The authors propose VLA-SCT, a lightweight, training-free self-correction framework that adds Trajectory Evaluation, Grasp Perturbation, and Termination Detection as an external control loop to existing VLA models. On LIBERO, VLA-SCT delivers state-of-the-art performance, achieving an average success rate of 81.55% with a 1.12× speedup over OpenVLA by improving both action precision and timely termination. This approach enhances robustness and deployment feasibility for VLA agents in complex, unstructured settings and suggests broad applicability across VLA architectures.

Abstract

While vision-language-action (VLA) models for embodied agents integrate perception, reasoning, and control, they remain constrained by two critical weaknesses: first, during grasping tasks, the action tokens generated by the language model often exhibit subtle spatial deviations from the target object, resulting in grasp failures; second, they lack the ability to reliably recognize task completion, which leads to redundant actions and frequent timeout errors. To address these challenges and enhance robustness, we propose a lightweight, training-free framework, VLA-SCT. This framework operates as a self-correcting control loop, combining data-driven action refinement with conditional logic for termination. Consequently, compared to baseline approaches, our method achieves consistent improvements across all datasets in the LIBERO benchmark, significantly increasing the success rate of fine manipulation tasks and ensuring accurate task completion, thereby promoting the deployment of more reliable VLA agents in complex, unstructured environments.
Paper Structure (10 sections, 7 equations, 3 figures, 2 tables)

This paper contains 10 sections, 7 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Trajectory Evaluation
  4. Grasp Perturbation
  5. Termination Detection
  6. Experiments
  7. Dataset and Metrics
  8. Implementation Details
  9. Experimental Results
  10. CONCLUSION

Figures (3)

  • Figure 1: The VLA-SCT framework addresses two fundamental challenges in VLAs. (a) For motion accuracy, it corrects a deviating trajectory. (b) For task termination, it prevents superfluous actions after task completion.
  • Figure 2: An overview of the VLA-SCT framework. Our method enhances a standard VLA model by introducing a conditional control loop for action self-correction and a visual judgment module for reliable task termination.
  • Figure 3: Sensitivity analysis of the trajectory quality threshold.