Safe Robotics Assembly through Reinforcement Learning

Researchers from Shanghai Jiao Tong University have made significant breakthroughs in robotics assembly using reinforcement learning, overcoming the challenges of repetitive and sensitive tasks. The team's visuo-tactile approach has achieved remarkable success in terminal assembly, reducing collisions and ensuring parts are handled with precision. The approach involves decomposing the assembly task into three phases, leveraging human demonstrations and interventions for safe training. Experimental results demonstrate the effectiveness of the method, achieving 100% successful insertions across various initial end-effector and grasp poses.

Key Takeaways:

• The proposed visuo-tactile assembly policy is robust to variations in grasp poses, minimizing collisions between the terminal head and terminal base.
• The approach involves three distinct phases: vision-guided model training, tactile-based grasp pose estimation, and visuo-tactile policy learning.
• The robot leverages human demonstrations and interventions to ensure safe training and prevent slippage or part breakage.
• Experimental results show that the proposed method achieves 100% successful insertions across 100 different initial end-effector and grasp poses.
• Imitation learning and online-RL policy yield only 9% and 0% successful insertions, respectively.
• The research was conducted by Yuchao Li, Ziqi Jin, Jin Liu, and Daolin Ma from the School of Ocean and Civil Engineering, Shanghai Jiao Tong University.
• The study aims to improve the efficiency and safety of industrial robotics assembly tasks.

Statistics:

• 100% successful insertions achieved by the proposed method across 100 different initial end-effector and grasp poses.
• 9% successful insertions achieved by imitation learning across 100 different initial end-effector and grasp poses.
• 0% successful insertions achieved by online-RL policy across 100 different initial end-effector and grasp poses.
• The research team used 100 different initial end-effector and grasp poses for testing the proposed method.

Sources:

• Visuo-tactile feedback policies for terminal assembly facilitated by reinforcement learning. Frontiers in Robotics and AI, 2025,12. (Frontiers in Robotics and AI - http://www.frontiersin.org/Robotics_and_AI)
• Shanghai Jiao Tong University Researchers Report on Findings in Robotics and Artificial Intelligence (Visuo-tactile feedback policies for terminal assembly facilitated by reinforcement learning). Journal of Engineering. November 3, 2025; p 3887