Show simple item record

dc.rights.licenseUsers may download and share copies with attribution in accordance with a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 License. No commercial use or derivatives are permitted without the explicit approval of the author.
dc.contributorTrinkle, Jeffrey C.
dc.contributorLyu, Siwei
dc.contributorCutler, Barbara M.
dc.contributorWen, John T.
dc.contributorPatterson, Stacy
dc.contributor.authorLi, Shuai
dc.date.accessioned2021-11-03T08:54:27Z
dc.date.available2021-11-03T08:54:27Z
dc.date.created2017-11-10T13:05:31Z
dc.date.issued2017-08
dc.identifier.urihttps://hdl.handle.net/20.500.13015/2076
dc.descriptionAugust 2017
dc.descriptionSchool of Science
dc.description.abstractIn this thesis, we focus on improving the perception capability for robots and addressing the problem of combining the perception capability with action planning and execution for robotic manipulation tasks. Our proposed approach combines Bayesian filtering methods with accurate models of multi-body dynamics for state estimation in the robotic manipulation tasks. In order to understand the design trade-offs of particle filter applications for the state estimation problems, we evalu- ate different particle filter modeling options in both simulation and physical exper- iments. We then propose a contact-based RBPF that samples the discrete contact states and updates the continuous state distribution through Kalman filters. Re- sults show that the contact-based RBPF is more effective and more efficient than the state of the art filters that sample the continuous state space. Finally, we apply reinforcement learning algorithms to learn policies for robotic manipulation tasks with a state space discretized using contact states. This discretized space learning is proven to be more effective than learning with continuous state space. We further propose to combine the learned policies with the contact-based RBPF for online action selection during robotic manipulation tasks.
dc.description.abstractThe robotic manipulation problem is very important in robotic research and appli- cations. In a typical robotic manipulation task, the robot needs to interact with certain objects to accomplish a goal, such as picking up a cup from a table. Without manipulation capabilities, robots will not be able to help humans with their daily tasks such as using tools to fix a broken car. Although the hardware of robots has been significantly improved, the ability to perceive the current state of a robotic manipulation task is still essential for a robot to fully utilize its hardware. However, less effort has been made to address the perception problem for robotic manipulation tasks.
dc.language.isoENG
dc.publisherRensselaer Polytechnic Institute, Troy, NY
dc.relation.ispartofRensselaer Theses and Dissertations Online Collection
dc.subjectComputer science
dc.titleContact-based state estimation and policy learning for robotic manipulation tasks
dc.typeElectronic thesis
dc.typeThesis
dc.digitool.pid178636
dc.digitool.pid178640
dc.digitool.pid178643
dc.rights.holderThis electronic version is a licensed copy owned by Rensselaer Polytechnic Institute, Troy, NY. Copyright of original work retained by author.
dc.description.degreePhD
dc.relation.departmentDept. of Computer Science


Files in this item

Thumbnail

This item appears in the following Collection(s)

Show simple item record