Improved deep Lagragian network-enabled momentum observer for collision detection during human-robot collaboration

Abstract

During human-robot collaboration (HRC), robots share workplaces with humans, and there may be frequent contact between them. It is crucial to be able to detect unexpected collisions in real-time so that appropriate safety measures should be taken to avoid injuries to humans and damage to robots. However, there are challenges with existing collision detection strategies, such as the additional costs incurred in deploying sensors in robots to implement pre-collision safety surveillance solutions or conducting complicated experiments to develop post-collision compliance solutions. To address these challenges, this paper presents a new momentum observer-based collision detection approach in which the external torques caused by collisions on robots can be efficiently identified. The approach involves integrating an improved deep Lagrangian network (DeLaN) to model robot dynamics without dynamic parameter identification experiments and prior knowledge of the robot’s physical and structural parameters. Another innovation of this approach is that a compensatory safety threshold is designed to enhance collision detection accuracy. Three robot datasets were used to train the improved DeLaN model. Simulation and real-world experiments were further carried out on the proposed approach to validate the effectiveness of the approach. Comparative experiments showed that the proposed approach outperformed other momentum observers in terms of both speed and efficiency. Moreover, experiments showed that the compensatory safety threshold proposed in this approach mitigated false positives caused by friction errors in robot joints to prevent the misdetection of collisions.