Abstract:
Numerous manipulation tasks, such as plug insertion and pipe assembly, demand an extremely high level of precision in pose estimation. Even minor errors, on the order of 2mm, can lead to task failure. While robots often rely on vision for object detection and localization, achieving consistent, high-precision localization using visual methods is not always feasible. This challenge is particularly pronounced in environments where working conditions like poor lighting, environmental occlusions, limited field of view, and lack of features, make vision challenging. When we as humans operate under such constraints we rely heavily on using tactile feedback to better perceive the environment and localize the object of interest sufficiently to complete the required tasks. In this thesis, we explore search-based planning techniques that would allow robots to do the same; and effectively utilize touch/contact modality to better perceive the environment, i.e., localize the object of interest, to the degree required for completing high-precision manipulation tasks.

The problem of planning for localization falls under the broader category of planning under uncertainty, typically formulated as a Partially Observable Markov Decision Process (POMDP), which is computationally expensive to solve. In the completed works of this thesis, we investigate the problem of localizing a target object using contact feedback to complete tasks under two different settings. We develop planning frameworks to solve the associated POMDPs, meeting the performance requirements of each setting.

The first framework is an experience-based preprocessing solution designed for semi-structured settings that require strong online performance. This means minimal online planning time while maintaining a high level of solution quality. We propose an experience-based POMDP solver that utilizes solutions of similar planning problems to speed up planning queries while maintaining strong bounds on solution quality. This solver is utilized to construct a database of solution policies offline which are queried online based on the problem encountered. The second framework is an online planning solution designed for less-structured domains (where preprocessing solutions are infeasible). Due to the need for solving problems online and the magnitude of uncertainty being larger in this setting, we propose a closed-loop planning and executing framework that utilizes a hierarchical representation of uncertainty. As the problems are solved online, this approach trades off solution quality to achieve reduced planning time. We demonstrate the effectiveness of both frameworks in a real-world plug insertion task in the presence of port pose uncertainty and a pipe assembly task in simulation.

For the remainder of this thesis, we will be exploring two broad directions. The first direction we propose to pursue is speeding up execution. The utilization of an enriched action and observation space has the potential to expedite task execution. Including impedance/admittance controller-based actions that enable the robot to glide along the surface of the target object (as opposed to probing the object) can help collect rich signals quickly leading to faster localization. Likewise, utilizing a denser observation space (like raw force-torque readings) can help localize the target object faster. However, denser observation signals often come with the downside of being highly noisy and continuous, which needs to be appropriately accounted for. The second proposed direction is to plan in the presence of unknown/unmodelled obstacles. There are often situations where the robot would need to manipulate in the presence of unmodelled obstacles. Localizing the target object under such conditions is particularly challenging as the robot would now have to discern between the interactions it makes with the target object and those occurring with the unmodelled obstructions making the planning problem challenging.

Thesis Committee:
Maxim Likhachev, Chair
Oliver Kroemer
Jeffrey Ichnowski
Mehmet Dogar, University of Leeds

More Information