{"id":{"repo_id":"uiuc","oai_identifier":"oai:www.ideals.illinois.edu:2142/121465"},"canonical_url":"https://search.dev.ndltd.org/etd/uiuc/oai:www.ideals.illinois.edu:2142/121465","repository":{"repo_id":"uiuc","name":"University of Illinois - Urbana-Champaign","base_url":"https://www.ideals.illinois.edu/oai-pmh"},"display":{"title":"Data-efficient learning for manipulation, locomotion, and information gathering involving granular media and deformable objects","abstract":"Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2023-12-04 without embargo terms","abstract_html":"Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2023-12-04 without embargo terms","abstract_has_math":false,"creators":["Zhu, Yifan"],"institution":"University of Illinois at Urbana-Champaign","degree_name":"Ph.D.","degree_level":"Dissertation","degree_discipline":"Computer Science","degree_department":null,"school":null,"contributors":["Hauser , Kris","Driggs-Campbell, Katie","Forsyth, David","Berenson, Dmitry","Goldberg, Ken"],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2023,"date_issued":"2023-08","date_published":"2023-08","updated_at":"2026-07-22T22:24:57Z","subjects":["Granular Media","Deformable Objects","Data-efficient Learning","Manipulation","Locomotion","Few-shot Learning"],"languages":["en","eng"],"rights":["2023 Yifan Zhu"],"rights_urls":[],"identifier_entries":[]},"links":{"outbound_url":"https://hdl.handle.net/2142/121465","outbound_label":"Handle","outbound_source":"dc:identifier"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor","label":"Contributor","values":["Hauser , Kris","Driggs-Campbell, Katie","Forsyth, David","Berenson, Dmitry","Goldberg, Ken"]},{"key":"dc:creator","label":"Author","values":["Zhu, Yifan"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date","label":"Dc Date","values":["2023-08","2023-07-10"]},{"key":"dc:type","label":"Dc Type","values":["text"]},{"key":"thesis:degree_discipline","label":"Discipline","values":["Computer Science"]},{"key":"thesis:degree_level","label":"Degree Level","values":["Dissertation"]},{"key":"thesis:degree_name","label":"Degree Name","values":["Ph.D."]},{"key":"thesis:institution_name","label":"Thesis Institution Name","values":["University of Illinois at Urbana-Champaign"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["Granular Media","Deformable Objects","Data-efficient Learning","Manipulation","Locomotion","Few-shot Learning"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["en","eng"]},{"key":"dc:rights","label":"Dc Rights","values":["2023 Yifan Zhu"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["https://hdl.handle.net/2142/121465"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2023-12-04 without embargo terms","The student, Yifan Zhu, accepted the attached license on 2023-07-06 at 13:15.","The student, Yifan Zhu, submitted this Dissertation for approval on 2023-07-06 at 13:25.","This Dissertation was approved for publication on 2023-07-10 at 16:32.","DSpace SAF Submission Ingestion Package generated from Vireo submission #19538 on 2023-12-04 at 17:00:51","Models of robots and how they contact the external world is traditionally built based on physics. However, such an approach is limited when the physics of certain phenomena are not well understood, when it is computationally prohibitive to solve for the equations, and when identifying equation parameters and solving conditions is challenging in the real world with partial and noisy observations. Recently, advancements in deep learning have provided a potential way to deal with this challenge, leveraging extremely flexible function approximators such as neural networks. However, the data required for many common robotics tasks could be prohibitive due to the complexity of the physics involved. This thesis aims to make progress toward addressing the issue of data efficiency for complex physics phenomena such as granular media, heterogeneous deformable objects, and acoustics of human bodies. To this end, this thesis adopts two main methodologies. First, a gray-box learning approach where learning is tightly integrated with physics, is employed to improve data efficiency. The core idea here is to decompose physics into parts that can be described by efficient analytical equations, and parts that are poorly understood or computationally heavy, which are learned from data. In this thesis, I will demonstrate different ways of combining knowledge of physics and learning to achieve data efficiency on multiple challenging problems. The second methodology aims to use meta-learning, or learning to learn, to extract useful prior knowledge from offline data on related tasks to accelerate online learning on novel tasks. I will demonstrate a novel meta-learning technique that enables a robot to use vision and very little online experience to achieve high-quality scooping actions on out-of-distribution granular terrains. We further show that these two methodologies can complement each other by demonstrating that the proposed meta-learning algorithm can improve gray-box learning for deformable objects. In addition to these two main methodologies, I also discuss my other relevant efforts in solving contact-rich robotics tasks, including automated excavation and manipulation in unstructured environments with an immersive, novice-friendly avatar robot that achieved 4-th place in the ANA XPRIZE Avatar Challenge."]},{"key":"dc:format","label":"Dc Format","values":["application/pdf"]},{"key":"dc:title","label":"Title","values":["Data-efficient learning for manipulation, locomotion, and information gathering involving granular media and deformable objects"]}]}],"canonical_facts":{"dc:contributor":["Hauser , Kris","Driggs-Campbell, Katie","Forsyth, David","Berenson, Dmitry","Goldberg, Ken"],"dc:creator":["Zhu, Yifan"],"dc:date":["2023-08","2023-07-10"],"dc:description":["Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2023-12-04 without embargo terms","The student, Yifan Zhu, accepted the attached license on 2023-07-06 at 13:15.","The student, Yifan Zhu, submitted this Dissertation for approval on 2023-07-06 at 13:25.","This Dissertation was approved for publication on 2023-07-10 at 16:32.","DSpace SAF Submission Ingestion Package generated from Vireo submission #19538 on 2023-12-04 at 17:00:51","Models of robots and how they contact the external world is traditionally built based on physics. However, such an approach is limited when the physics of certain phenomena are not well understood, when it is computationally prohibitive to solve for the equations, and when identifying equation parameters and solving conditions is challenging in the real world with partial and noisy observations. Recently, advancements in deep learning have provided a potential way to deal with this challenge, leveraging extremely flexible function approximators such as neural networks. However, the data required for many common robotics tasks could be prohibitive due to the complexity of the physics involved. This thesis aims to make progress toward addressing the issue of data efficiency for complex physics phenomena such as granular media, heterogeneous deformable objects, and acoustics of human bodies. To this end, this thesis adopts two main methodologies. First, a gray-box learning approach where learning is tightly integrated with physics, is employed to improve data efficiency. The core idea here is to decompose physics into parts that can be described by efficient analytical equations, and parts that are poorly understood or computationally heavy, which are learned from data. In this thesis, I will demonstrate different ways of combining knowledge of physics and learning to achieve data efficiency on multiple challenging problems. The second methodology aims to use meta-learning, or learning to learn, to extract useful prior knowledge from offline data on related tasks to accelerate online learning on novel tasks. I will demonstrate a novel meta-learning technique that enables a robot to use vision and very little online experience to achieve high-quality scooping actions on out-of-distribution granular terrains. We further show that these two methodologies can complement each other by demonstrating that the proposed meta-learning algorithm can improve gray-box learning for deformable objects. In addition to these two main methodologies, I also discuss my other relevant efforts in solving contact-rich robotics tasks, including automated excavation and manipulation in unstructured environments with an immersive, novice-friendly avatar robot that achieved 4-th place in the ANA XPRIZE Avatar Challenge."],"dc:format":["application/pdf"],"dc:identifier":["https://hdl.handle.net/2142/121465"],"dc:language":["en","eng"],"dc:rights":["2023 Yifan Zhu"],"dc:subject":["Granular Media","Deformable Objects","Data-efficient Learning","Manipulation","Locomotion","Few-shot Learning"],"dc:title":["Data-efficient learning for manipulation, locomotion, and information gathering involving granular media and deformable objects"],"dc:type":["text"],"thesis:degree_discipline":["Computer Science"],"thesis:degree_level":["Dissertation"],"thesis:degree_name":["Ph.D."],"thesis:institution_name":["University of Illinois at Urbana-Champaign"]},"updated_at":"2026-07-22T22:24:57Z"}