An Optimization-Based Human Behavior Modeling and Prediction for Human-Robot Collaborative Disassembly

Abstract

To achieve a safe and seamless human-robot collaboration in intelligent remanufacturing, robot agents should be able to understand human behaviors, predict human future motion, and incorporate motion prediction into their planning process. While most existing human prediction algorithms suffer from poor generalization and huge training data requirements, this paper models the human agent as a rational model seeking to minimize an unknown cost function along the motion trajectory. With such modeling, we design a set of features, such as collision avoidance, maintaining comfort during the motion, and reaching the goal point without too much detour, that could capture human intents during HRC. Maximum-Entropy inverse reinforcement learning is then leveraged to learn the underlying cost function from noisy human demonstrations. The human motion prediction is obtained by solving an optimization problem with a learned cost function. We particularly build an HRC dataset for human-robot-collaborative disassembly tasks and applied the proposed algorithm to this new dataset. Experimental studies are extensively conducted to validate our human motion prediction model.