Safe Human-Robot Interaction Through Machine Learning and Reachability Analysis

• Applicant: Professor Dr.-Ing. Matthias Althoff
• Subject Area: Automation, Mechatronics, Control Systems, Intelligent Technical Systems, Robotics
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 548385598

Many processes cannot be automated by robots because the required cognitive capabilities are too high. A promising solution to this challenge is the integration of collaborative robots (aka cobots) to support human workers in physically demanding tasks. To successfully integrate cobots, they must be safe despite relatively high velocities and must be capable of solving a wide range of collaborative tasks, as most of these tasks cannot be programmed like traditional manufacturing processes. To address the above-mentioned challenges, we will develop a flexible framework for the safety of AI-controlled robots. By leveraging reinforcement learning and reachability analysis, our proposed safety framework aims to make formally correct decisions that not only ensure human safety but also improve task efficiency and alignment with human preferences compared to traditional planning algorithms. We propose a variety of safety modes, between which our safety framework can switch based on the current requirements on-the-fly. By offering multiple safety modes, our framework can accommodate various human preferences and operational requirements, resulting in a more versatile and adaptable solution for safe human-robot interaction. In addition, to make reinforcement learning more competitive for human-robot interaction, we will develop a novel shadow mode in which the reinforcement learning agent trains while the system is operated by a conventional controller, which instantaneously performs well as it does not have to be trained. Once the reinforcement learning agent is confident enough to be better than the conventional controller in certain operating regions, those regions will be handed over to the reinforcement learning agent. We will demonstrate our approach not only on many benchmarks, but also on a Schunk LWA 4P robot and a modular robot from RobCo.

• DFG Programme: Research Grants