Toward Reinforcement Learning for Selection and Parameterization of Appropriate Prevention Measures for Collaborative Robotic Applications

Abstract

The deployment of collaborative applications has emerged as a key trend in the manufacturing industry, driven by the synergy of human intelligence and adaptability with robots' endurance, strength, and precision. However, this closer integration of humans with robots heightens the risk of injuries. To prevent these risks, the risk reduction process is performed during design phase of collaborative workstations, ensuring that safety criteria required by current standards are met. In this work, we formulate the risk reduction process as a sequential decision-making problem and propose a Reinforcement Learning-based approach. In a simulation environment, the RL agent learns to select and parameterize prevention measures to reduce collision and crushing risks of collaborative applications. The considered prevention measures include physical barriers, laser safety barriers, and Tool Center Point (TCP) speed monitoring zones parameterized by their geometric and temporal activation interval parameters. The RL agent is trained through trial-and-error interactions within the simulated environment, aiming to maximize a reward function that prioritizes risk reduction. Secondary objectives include minimizing cycle time and reducing workstation footprint, allowing the system to derive optimized solutions that satisfy safety while balancing cost and productivity constraints. A feasibility demonstrator is presented through a use case in which the RL agent proposes prevention measures to reduce identified hazards for a single collaborative application, enabling faster human-centered design of safe collaborative workstations.

Keywords: Reinforcement Learning, Robotics Safety, Human-Robot Collaboration, Risk Reduction

DOI: 10.54941/ahfe1008091