A Dual-Functional Push-Grasping Strategy for Goal-Agnostic and Goal-Oriented Tasks Based on Deep Reinforcement Learning

In the field of robotic grasping, tasks can be categorized into two primary types: goal-agnostic tasks and goal-oriented tasks. The former aims to retrieve all objects within the workspace, whereas the latter focuses on grasping a specific predefined goal object. Current grasping techniques often excel at one type of task while underperforming at the other. To address this, an innovative dual-functional push-grasp synergy strategy has been proposed, which integrates pushing and grasping actions to efficiently handle both types of tasks efficiently. This strategy relies on a dual-functional network that utilizes visual inputs to generate dense pixel-wise Q-value maps for pushing and grasping actions, thereby expanding the pool of available action samples. The goal-agnostic task is conceptualized as a collection of multiple goal-oriented tasks, and a hierarchical reinforcement learning framework has been designed to coordinate task execution. To simplify the training process, a two-stage training approach has been adopted, optimizing separately for each task type. In goal-oriented tasks, the effectiveness of actions and the efficiency of network training are enhanced by incorporating constraints on irrational actions and adjusting the reward function based on changes in free space. Pre-training is performed in a simulated environment, after which the model is deployed in physical scenarios without further calibration. Through evaluation, it was established that the proposed approach yields substantially higher rates of task completion and successful grasps than existing methods.