Vision-Based Contact Wrench Estimation in Human-Robot Interaction

Abstract

With the rapid integration of robotics across diverse sectors, human interaction with these technologies is becoming inevitable. Ensuring safety is increasingly crucial to prevent injuries and maintain effective interactions. Accurate force estimation enables robots to sense contact forces and respond appropriately. This paper presents a vision-based estimation method for multi-contact physical human-robot interaction. Utilizing an RGB-D sensor, it detects 3D hand positions to identify contact points and employs a generalized momentum observer to distinguish joint torques from external wrenches. A long short-term memory network compensates for uncertainties arising from unmodelled dynamics. Addressing challenges like wrench null space and Jacobian singularities, the approach identifies computable external wrench components. The method achieves a 0.9 N estimation error in complex, multi-contact interactions, enhancing safety and responsiveness. Key contributions include a novel wrench identification method leveraging robot configuration and contact points, derived from a vision-based system, to enhance real-time estimation.