engleski

Human action and motion prediction in industrial human-robot shared environments using probabilistic decision-making methods

As robots are progressing towards being ubiquitous and an indispensable part of our everyday environments efficient and safe collaboration and cohabitation become imperative. Given that, such environments could benefit greatly from accurate human action and motion prediction. In addition to being accurate, human action prediction should be computationally efficient, in order to ensure a timely reaction, and capable of dealing with changing environments, since unstructured interaction and collaboration with humans usually do not assume static conditions. In this thesis, we focus on probabilistic decision-making models for human action and motion prediction in industrial human-robot shared environments. We firstly introduce a framework for human intention recognition in the robotized ware- house environment. This framework is based on Markov Decision Process action validation and the Hidden Markov Model intention recognition module. The warehouse floormap is searched to find optimal paths towards potential goals using Generalized Voronoi Diagrams and D∗ graph search algorithms. We continued by utilizing this framework for precise human motion prediction that served as input for the human-aware planning algorithm. The goal of human-aware planning is to reroute robots in the worker’s way thus improving efficiency and retaining safety. We ran multiple experiments to test the proposed algorithms: in a real-world laboratory warehouse with a worker, wearing augmented reality glasses, a virtual reality warehouse twin as well an in-house developed warehouse simulator. Finally, we utilized Long Short-Term Memory networks to predict the next object human is going to pick in a collaborative environment. In order to reduce execution time, we crafted two dimensionality reduction methods. The first one is a feature selection method that relies on signal correlation and individual merit while the second one is a feature extraction method based on the autoencoder model. Heavy emphasis was put on the best predictor for human action prediction, the eye gaze, and an effort was made to estimate it using Multilayer Perceptron architecture. We used motion capture data from the publicly available dataset as well as on a smaller in-house recorded dataset.

human action prediction ; human motion prediction ; Markov decision process ; hidden Markov model ; probabilistic decision-making models ; recurrent neural networks ; long short-term memory networks ; feature dimensionality reduction ; autoencoders ; collaborative environments

nije evidentirano