Researchers at Shandong University have developed a robotic grasping method that coordinates pushing and grasping skills to handle complex household scenes. The approach, detailed in the journal Robot, models the task as a Markov decision process and introduces two neural networks: a pushing-grasping skill confidence evaluation network (PGCE-Net) and a clutter level evaluation network (CE-Net). PGCE-Net outputs pixel-level confidence maps for grasping, pushing, and pushing distance, enabling the robot to select the most promising action. CE-Net assesses scene clutter into three levels (0, 1, 2) and adjusts the robot's action strategy accordingly—direct grasping for low clutter, more pushing for high clutter. The method also designs tailored reward functions to enhance both skills. Experiments in simulation (CoppeliaSim) and real environments validated the approach, showing improved grasping success rates compared to existing methods. This work addresses challenges in unstructured home settings where objects vary in shape, color, and arrangement, offering a practical solution for service robots and automated picking systems.
Key takeaways
- PGCE-Net enables pixel-level evaluation of pushing and grasping actions, with 16 directions and pushing distances from 0.05 m to 0.15 m.
- CE-Net classifies scene clutter into three levels, guiding the robot to prioritize pushing or grasping based on object density.
- The method was validated in both simulated and real-world robotic experiments, demonstrating feasibility and effectiveness.
For system integrators and automation engineers, this research highlights the potential of combining deep reinforcement learning with skill collaboration to improve robotic manipulation in unstructured environments, a key step toward more autonomous service and industrial robots.
Source: 《机器人》期刊 (robot.sia.cn) · Published 2026-01-13 · “基于推抓技能协同的家庭复杂场景下的物品抓取方法”