Dynamic Closed-Loop Grasping Via an Enhanced Thinkgrasp Framework

Abstract

To address the problems of robotic target part grasping in cluttered and occluded environments, such as poor integration of semantic decision-making and execution modules, limited adaptability of single-step planning to dynamic occlusions, and insufficient stability in target recognition, this paper proposes a vision-language guided closed-loop grasping method based on the ThinkGrasp framework. This paper uses the Kimi-K2.6 multimodal model for semantic decision-making and applies a $\mathbf{3} \boldsymbol{\times} \mathbf{3}$ grid constraint to reduce localisation errors. It integrates LangSAM and VLPart for adaptive target segmentation and uses GraspNet to generate and filter 6-DoF grasp poses. A closed-loop iterative mechanism is also used to gradually eliminate occlusions through several rounds of perception and action sequences to achieve continuous grasping in dynamic environments. Two types of test scenarios have been built in the PyBullet simulation platform: general clutter and heavy occlusion, and several popular algorithms have been tested comparatively. Modules ablation studies were conducted to determine the effect of modules. Based on the experiments, our method has a grasp success rate of nearly 100% for general scenes and an average success rate of 81.1% in heavy occlusion situations, which is an improvement of 2.2 percentage points over the original framework. In addition, the mean number of grasp steps has decreased by 11.6%, and thus the whole system is more stable and efficient.