{"id":{"repo_id":"mit","oai_identifier":"oai:dspace.mit.edu:1721.1/153869"},"canonical_url":"https://search.dev.ndltd.org/etd/mit/oai:dspace.mit.edu:1721.1/153869","repository":{"repo_id":"mit","name":"MIT","base_url":"https://dspace.mit.edu/oai/request"},"display":{"title":"Learning Compositional Abstract Models Incrementally for Efficient Bilevel Task and Motion Planning","abstract":"In robotic domains featuring continuous state and action spaces, planning in long-horizon task is fundamentally hard, even when the transition model is deterministic and known. One way to alleviate this challenge is to perform bilevel planning with abstractions, where a high-level search for abstract plans is used to guide planning in the original transition space. In this thesis, we propose an algorithm for learning predicates from demonstrations, eliminating the need for manually specified state abstractions. Our key idea is to learn predicates by optimizing a surrogate objective that is tractable but faithful to our real efficient-planning objective. We use this surrogate objective in a hill-climbing search over predicate sets drawn from a grammar, which we call predicate invention. However, our research highlights another limitation in current symbolic operator learning techniques. They often fall short in robotics scenarios where the robot’s actions result in numerous inconsequential alterations to the abstract state. This limitation arises mainly because these techniques aim to precisely predict every observed change in that state, and as the execution horizon grows longer so does the built up complexity of the predictions. In this thesis, we study this separately and introduce an innovative method where the operators are induced to selectively predict by focusing solely on changes crucial for abstract planning to meet specific subgoals, which we call our operator learning procedure. Our contributions include: a predicate invention procedure based on a hill-climbing search over predicate sets, and a planning-driven operator learning objective based on a hill-climbing search algorithm that only model changes necessary for abstract planning and preserve compositionality of operators. We evaluate learning predicates and operators across a few toy environments and dozens of tasks from the demanding BEHAVIOR-100 benchmark.","abstract_html":"In robotic domains featuring continuous state and action spaces, planning in long-horizon task is fundamentally hard, even when the transition model is deterministic and known. One way to alleviate this challenge is to perform bilevel planning with abstractions, where a high-level search for abstract plans is used to guide planning in the original transition space. In this thesis, we propose an algorithm for learning predicates from demonstrations, eliminating the need for manually specified state abstractions. Our key idea is to learn predicates by optimizing a surrogate objective that is tractable but faithful to our real efficient-planning objective. We use this surrogate objective in a hill-climbing search over predicate sets drawn from a grammar, which we call predicate invention. However, our research highlights another limitation in current symbolic operator learning techniques. They often fall short in robotics scenarios where the robot’s actions result in numerous inconsequential alterations to the abstract state. This limitation arises mainly because these techniques aim to precisely predict every observed change in that state, and as the execution horizon grows longer so does the built up complexity of the predictions. In this thesis, we study this separately and introduce an innovative method where the operators are induced to selectively predict by focusing solely on changes crucial for abstract planning to meet specific subgoals, which we call our operator learning procedure. Our contributions include: a predicate invention procedure based on a hill-climbing search over predicate sets, and a planning-driven operator learning objective based on a hill-climbing search algorithm that only model changes necessary for abstract planning and preserve compositionality of operators. We evaluate learning predicates and operators across a few toy environments and dozens of tasks from the demanding BEHAVIOR-100 benchmark.","abstract_has_math":false,"creators":["McClinton III, Willie B."],"institution":"Massachusetts Institute of Technology","degree_name":"Master","degree_level":null,"degree_discipline":null,"degree_department":"Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science","school":null,"contributors":[],"advisors":["Kaelbling, Leslie Pack","Lozano-Pérez, Tomás"],"committee_chairs":[],"committee_members":[],"year":2024,"date_issued":"2024-02","date_published":"2024-02","updated_at":"2026-07-22T22:21:05Z","subjects":[],"languages":[],"rights":["In Copyright - Educational Use Permitted","Copyright retained by author(s)"],"rights_urls":["https://rightsstatements.org/page/InC-EDU/1.0/"],"identifier_entries":[]},"links":{"outbound_url":"https://hdl.handle.net/1721.1/153869","outbound_label":"Handle","outbound_source":"dc:identifier.uri"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor.advisor","label":"Advisor","values":["Kaelbling, Leslie Pack","Lozano-Pérez, Tomás"]},{"key":"dc:contributor.department","label":"Department","values":["Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science"]},{"key":"dc:creator","label":"Author","values":["McClinton III, Willie B."]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.accessioned","label":"Dc Date Accessioned","values":["2024-03-21T19:12:11Z"]},{"key":"dc:date.available","label":"Dc Date Available","values":["2024-03-21T19:12:11Z"]},{"key":"dc:date.issued","label":"Date","values":["2024-02"]},{"key":"dc:publisher","label":"Institution","values":["Massachusetts Institute of Technology"]},{"key":"dc:type","label":"Dc Type","values":["Thesis"]},{"key":"thesis:degree_name","label":"Degree Name","values":["Master","Master of Science in Electrical Engineering and Computer Science"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:rights","label":"Dc Rights","values":["In Copyright - Educational Use Permitted","Copyright retained by author(s)"]},{"key":"dc:rights.uri","label":"Rights URI","values":["https://rightsstatements.org/page/InC-EDU/1.0/"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://hdl.handle.net/1721.1/153869"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.abstract","label":"Abstract","values":["In robotic domains featuring continuous state and action spaces, planning in long-horizon task is fundamentally hard, even when the transition model is deterministic and known. One way to alleviate this challenge is to perform bilevel planning with abstractions, where a high-level search for abstract plans is used to guide planning in the original transition space. In this thesis, we propose an algorithm for learning predicates from demonstrations, eliminating the need for manually specified state abstractions. Our key idea is to learn predicates by optimizing a surrogate objective that is tractable but faithful to our real efficient-planning objective. We use this surrogate objective in a hill-climbing search over predicate sets drawn from a grammar, which we call predicate invention. However, our research highlights another limitation in current symbolic operator learning techniques. They often fall short in robotics scenarios where the robot’s actions result in numerous inconsequential alterations to the abstract state. This limitation arises mainly because these techniques aim to precisely predict every observed change in that state, and as the execution horizon grows longer so does the built up complexity of the predictions. In this thesis, we study this separately and introduce an innovative method where the operators are induced to selectively predict by focusing solely on changes crucial for abstract planning to meet specific subgoals, which we call our operator learning procedure. Our contributions include: a predicate invention procedure based on a hill-climbing search over predicate sets, and a planning-driven operator learning objective based on a hill-climbing search algorithm that only model changes necessary for abstract planning and preserve compositionality of operators. We evaluate learning predicates and operators across a few toy environments and dozens of tasks from the demanding BEHAVIOR-100 benchmark."]},{"key":"dc:description.degree","label":"Dc Description Degree","values":["S.M."]},{"key":"dc:title","label":"Title","values":["Learning Compositional Abstract Models Incrementally for Efficient Bilevel Task and Motion Planning"]}]}],"canonical_facts":{"dc:contributor.advisor":["Kaelbling, Leslie Pack","Lozano-Pérez, Tomás"],"dc:contributor.department":["Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science"],"dc:creator":["McClinton III, Willie B."],"dc:date.accessioned":["2024-03-21T19:12:11Z"],"dc:date.available":["2024-03-21T19:12:11Z"],"dc:date.issued":["2024-02"],"dc:description.abstract":["In robotic domains featuring continuous state and action spaces, planning in long-horizon task is fundamentally hard, even when the transition model is deterministic and known. One way to alleviate this challenge is to perform bilevel planning with abstractions, where a high-level search for abstract plans is used to guide planning in the original transition space. In this thesis, we propose an algorithm for learning predicates from demonstrations, eliminating the need for manually specified state abstractions. Our key idea is to learn predicates by optimizing a surrogate objective that is tractable but faithful to our real efficient-planning objective. We use this surrogate objective in a hill-climbing search over predicate sets drawn from a grammar, which we call predicate invention. However, our research highlights another limitation in current symbolic operator learning techniques. They often fall short in robotics scenarios where the robot’s actions result in numerous inconsequential alterations to the abstract state. This limitation arises mainly because these techniques aim to precisely predict every observed change in that state, and as the execution horizon grows longer so does the built up complexity of the predictions. In this thesis, we study this separately and introduce an innovative method where the operators are induced to selectively predict by focusing solely on changes crucial for abstract planning to meet specific subgoals, which we call our operator learning procedure. Our contributions include: a predicate invention procedure based on a hill-climbing search over predicate sets, and a planning-driven operator learning objective based on a hill-climbing search algorithm that only model changes necessary for abstract planning and preserve compositionality of operators. We evaluate learning predicates and operators across a few toy environments and dozens of tasks from the demanding BEHAVIOR-100 benchmark."],"dc:description.degree":["S.M."],"dc:identifier.uri":["https://hdl.handle.net/1721.1/153869"],"dc:publisher":["Massachusetts Institute of Technology"],"dc:rights":["In Copyright - Educational Use Permitted","Copyright retained by author(s)"],"dc:rights.uri":["https://rightsstatements.org/page/InC-EDU/1.0/"],"dc:title":["Learning Compositional Abstract Models Incrementally for Efficient Bilevel Task and Motion Planning"],"dc:type":["Thesis"],"thesis:degree_name":["Master","Master of Science in Electrical Engineering and Computer Science"]},"updated_at":"2026-07-22T22:21:05Z"}