Deep Reinforcement Learning for Adaptive Robotic Grasping and Post-Grasp Manipulation in Simulated Dynamic Environments

This article presents a deep reinforcement learning (DRL) approach for adaptive robotic grasping in dynamic environments. We developed UR5GraspingEnv, a PyBullet-based simulation environment integrated with OpenAI Gym, to train a UR5 robotic arm with a Robotiq 2F-85 gripper. Soft Actor-Critic (SAC) and Proximal Policy Optimization (PPO) were implemented to learn robust grasping policies for randomly positioned objects. A tailored reward function, combining distance penalties, grasp, and pose rewards, optimizes grasping and post-grasping tasks, enhanced by domain randomization. SAC achieves an 87% grasp success rate and 75% post-grasp success, outperforming PPO 82% and 68%, with stable convergence over 100,000 timesteps. The system addresses post-grasping manipulation and sim-to-real transfer challenges, advancing industrial and assistive applications. Results demonstrate the feasibility of learning stable and goal-driven policies for single-arm robotic manipulation using minimal supervision. Both PPO and SAC yield competitive performance, with SAC exhibiting superior adaptability in cluttered or edge cases. These findings suggest that DRL, when carefully designed and monitored, can support scalable learning in manipulation tasks.