Planning under uncertainty in robotics

This thesis experimentally addresses the issue of planning under uncertainty in robotics, with reference to the Partially Observable Markov Decision Process (POMDP) framework. POMDP algorithms have been successfully used in many real-world applications, but they are sometimes avoided in robotics, especially when large state spaces and strict, short time constraints are involved. For this reason, the aim of this study is to test existing POMDP algorithms in large domains, with very short time limits. The thesis first examines the main sources of robots’ uncertainty about the world, thus motivating the further study of the POMDP framework. Indeed, it turns out that considering uncertainty during planning in robotics can be sometimes beneficial. Secondly, three approximate, online algorithms (POMCP, DESPOT and PGI) are selected given the challenges specific to a robotic setting, and the properties of the chosen algorithms are outlined. Finally, test frameworks and benchmark problems, when not provided by the authors, are implemented and used to test and compare the performance of the selected algorithms in large domains characterised by different sizes of state, observation and action spaces, with strict, short time limits. The results of the experiments show that smart prior knowledge greatly improves the performance of the algorithms and that the selected POMDP algorithms generally scale to large state and observation spaces better than to large action spaces. Moreover, the results with preferred actions show that both DESPOT and POMCP achieved high performances for relatively long time limits, while DESPOT performed the best for shorter ones, apart from one exception. Finally, the results with legal actions show that POMCP achieved the best performance in all the problems, except the smallest one, where PGI obtained similar results for short time limits, and both PGI and DESPOT outperformed POMCP with 1s of time limit