Searching for objects is prevalent in many settings, such as pharmacies, supermarkets, recycling plants and warehouses. Many approaches exist to plan such tasks, including heuristics, look-ahead planners and end-to-end machine learning. Currently, most approaches are specialised for either shelf or table environments, limiting their transferability to different environments. The transfer of approaches between shelf and table environments is assessed in this work. This is done by comparing two state-of-the-art policies designed for shelves in table environments. The policies are compared on their success rate, number of actions, and compute time. Distribution Area Reduction for Stacked Scenes (DARSS) and Monte Carlo Tree Searches for Stacked Scenes (MCTSSS) were selected as two state-of-the-art object search policies. DARSS is a heuristic greedy policy, whereas MCTSSS is a tree search-based look-ahead policy. The policies are tasked with searching for a known target object amongst stacks of unknown objects. Actions can be taken to move objects to uncover the target and complete the task. Both policies were compared in scenes with varying numbers of objects and stacks in the environment. MCTSSS was also compared with a different number of iterations and tree depth. The results of both policies in table environments were compared using Bayesian estimation and effect size calculations. Overall, MCTSSS, with a low number of iterations, performed the best, with a task completion rate of 87%. DARSS also performed well, with a task completion rate of 79%, and required an order of magnitude less time to compute. In contrast, MCTSSS did not perform well with a higher number of iterations. The number of objects and stacks was a significant factor for DARSS. This was not observed in the case of MCTSSS, suggesting that MCTSSS is more invariant to these changes. The number of actions taken was compared between the policies, but the differences were not significant. Overall, this suggests that DARSS and MCTSSS, with a low number of iterations, can transfer to table environments. Further studies are required to understand the limitations of these policies in real-world developments.
Date of Award | 31 Dec 2024 |
---|
Original language | English |
---|
Awarding Institution | - The University of Manchester
|
---|
Supervisor | Bruno Adorno (Supervisor) & Simon Watson (Supervisor) |
---|
- Object Search
- Bayesian Analysis
- robotic manipulator
- POMDP
- Robotics
- Bayesian
Understanding the Generalizability of Shelf Object Search Policies in Top-Down Access Environments
Basit, A. (Author). 31 Dec 2024
Student thesis: Master of Philosophy