Search

Autonomous mobile robots employed in industrial applications often operate in complex and uncertain environments. In this paper we propose an approach based on an extension of Partially Observable Monte Carlo Planning (POMCP) for robot velocity regulation in industrial-like environments characterized by uncertain motion difficulties. The velocity selected by POMCP is used by a standard engine controller which deals with path planning. This two-layer approach allows POMCP to exploit prior knowledge on the relationships between task similarities to improve performance in terms of time spent to traverse a path with obstacles. We also propose three measures to support human-understanding of the strategy used by POMCP to improve the performance. The overall architecture is tested on a Turtlebot3 in two environments, a rectangular path and a realistic production line in a research lab. Tests performed on a C++ simulator confirm the capability of the proposed approach to profitably use prior knowledge, achieving a performance improvement from 0.7% to 3.1% depending on the complexity of the path. Experiments on a Unity simulator show that the proposed two-layer approach outperforms also single-layer approaches based only on the engine controller (i.e., without the POMCP layer). In this case the performance improvement is up to 37% comparing to a state-of-the-art deep reinforcement learning engine controller, and up to 51% comparing to the standard ROS engine controller. Finally, experiments in a real-world testing arena confirm the possibility to run the approach on real robots. [ABSTRACT FROM AUTHOR]

Autonomous surface vessels are becoming increasingly important for water monitoring. Their aim is to navigate rivers and lakes with limited intervention of human operators, to collect real-time data about water parameters. To reach this goal, these intelligent systems must interact with the environment and act according to the situations they face. In this work we propose a framework based on the integration of recent time-series clustering/segmentation methods and cluster validity indices, for detecting, modeling and evaluating aquatic drone states. The approach is completely data-driven and unsupervised. It takes unlabeled multivariate time series of sensor traces and returns both a set of statistically significant state-models (generated by different mathematical approaches) and a related segmentation of the dataset. We test the approach on a real dataset containing data of six campaigns, two in rivers and four in lakes, in different countries for about 5.6 h of navigation. Results show that the methodology is able to recognize known states and to discover unknown states, enabling novelty detection. The approach is therefore an easy-to-use tool for discovering and interpreting significant states in sensor data, that enables improved data analysis and drone autonomy. [ABSTRACT FROM AUTHOR]