Compositional reasoning in robot learning
Abstract/Contents
- Abstract
- To carry out diverse tasks in everyday human environments, future robots must generalize beyond the knowledge they are equipped with. However, despite recent advances in "end-to-end" deep learning, today's robot learning methods are still limited to specializing in one task at a time. At the same time, humans perform everyday tasks with ease. But instead of learning each task in silos, we distill reusable abstractions from our daily experiences and solve new tasks by composing known building blocks. Such compositional reasoning capability is crucial for developing future robots that are both competent and versatile. This two-part thesis presents a spectrum of techniques for building compositional generalization capabilities into robot learning systems. Part I starts by drawing prominent ideas from classical structured approaches in AI such as program induction and graphical models and design algorithms with strong representational priors. Neural Task Programming (NTP) represents task-subtask hierarchies using modular neural programs. NTP is designed to simultaneously exploit the representational capacity of neural networks for handling unstructured input and the compositional nature of programs for few-shot generalization. Similarly, Neural Task Graphs (NTGs) represent manipulation skills as nodes in a graph and leverage structures in skill preconditions for generalization. Regression Planning Network (RPN) further dissects the goal specification by grounding symbolic goals with object-centric representation and achieve zero-shot generalization to new task goals. To bring compositional reasoning closer to real-world settings, Part II of this thesis introduces our recent efforts to relax the strong representational prior assumptions: learning from unstructured video demonstrations and learning through trial-and-error. We first describe a framework for simultaneously learning complex visuomotor skills and discovering implicit compositional structures from human demonstrations. Experiment shows the approach allows a physical robot to perform long-horizon manipulation in a kitchen setup. Then we further relax the assumption of having demonstration data and enable a robot to learn compact planning representations by characterizing the pre- and post-condition of motion primitives through active interaction.
Description
Type of resource | text |
---|---|
Form | electronic resource; remote; computer; online resource |
Extent | 1 online resource. |
Place | California |
Place | [Stanford, California] |
Publisher | [Stanford University] |
Copyright date | 2021; ©2021 |
Publication date | 2021; 2021 |
Issuance | monographic |
Language | English |
Creators/Contributors
Author | Xu, Danfei |
---|---|
Degree supervisor | Li, Fei Fei, 1976- |
Degree supervisor | Savarese, Silvio |
Thesis advisor | Li, Fei Fei, 1976- |
Thesis advisor | Savarese, Silvio |
Thesis advisor | Bohg, Jeannette, 1981- |
Thesis advisor | Sadigh, Dorsa |
Degree committee member | Bohg, Jeannette, 1981- |
Degree committee member | Sadigh, Dorsa |
Associated with | Stanford University, Computer Science Department |
Subjects
Genre | Theses |
---|---|
Genre | Text |
Bibliographic information
Statement of responsibility | Danfei Xu. |
---|---|
Note | Submitted to the Computer Science Department. |
Thesis | Thesis Ph.D. Stanford University 2021. |
Location | https://purl.stanford.edu/ys432pr6718 |
Access conditions
- Copyright
- © 2021 by Danfei Xu
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...