Reinforcement Learning-based Pushing-Grasping Tasks with Grasp Success Prediction

Robotic grasping in cluttered environments presents significant challenges due to occlusions and physical interactions that hinder successful object retrieval. Traditional grasping strategies often fail in densely cluttered scenes, where direct grasp attempts lead to frequent collisions or inaccessible target objects. To address this limitation, we propose an RL pushing-grasping framework that incorporates a grasp success prediction (GSP) model to guide pre-grasp manipulation. An autoencoder compresses RGB-D data and object masks into a latent representation for a predictor that estimates grasp success. When this estimate falls below a threshold, an SAC-trained agent first chooses to push to improve the object’s pose and then executes the grasp. The proposed method is evaluated in randomised cluttered environments within the PyBullet simulation platform. Experimental results demonstrate that our approach significantly improves grasping performance, achieving an 81% compared to 56% for a heuristic-based baseline, showing that it selects effective movements for rearranging the scene. These results show that predicting grasp success is important for deciding how to prepare the scene, helping the robot retrieve objects more efficiently and reliably in cluttered environments.