Learning, inference, and control for construction robots and spatiotemporal processes

Maske, Harshal

Permalink

https://hdl.handle.net/2142/101651 Copy

Description

Title

Learning, inference, and control for construction robots and spatiotemporal processes

Author(s)

Maske, Harshal

Issue Date

2018-06-15

Director of Research (if dissertation) or Advisor (if thesis)

Chowdhary, Girish

Doctoral Committee Chair(s)

Chowdhary, Girish

Committee Member(s)

• Bretl, Timothy
• Grift, Tony E.
• Pagilla, Prabhakar

Department of Study

Engineering Administration

Discipline

Agricultural & Biological Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Instructional Policy
• Robot Learning
• Shared control
• Kernel Observer

Abstract

Expert operators of real world robots, especially constructions robots, develop expertise from years of training and experience. In the absence of such experts, these robots are operated by novice operators, and this adversely affects the productivity. On the other hand, safety concerns and the nature of the operating environment limits the possibility of automating these robots. This thesis proposes a solution by considering a problem setting in which the robot learns a policy from experts to train novice human operators. Formally, this is posed as the problem of learning instructional policy from demonstration, that maps the state of the robot to an instruction for a human operator. Existing methods learn policy from demonstration, however such policies do not relate to the human operator's action space and hence cannot be used to generate instructions for novice operators. We introduce action primitives that address this challenge of mapping continuous state action trajectories to human parse-able and executable instructions. Construction tasks are complex as they consist of several subtasks with stochastic transitions. For such tasks, existing approaches learn policy for component subtask and then rely either on predefined decomposition or heuristics to generate policy for the entire task. To overcome this limitation, and to generate instructions for an entire construction task, this thesis proposes learning of a structured probabilistic model for instructional policy. This model utilizes hierarchy of Markov chains that incrementally captures the number of subtasks as well as their transitions. Switching between the subtasks is inferred using a likelihood rate based inference approach proposed in this thesis. Instructional policy model is tested based on a controlled group study involving 113 participants, who learn to perform the truck loading task on a hydraulic actuated scaled excavator robot. Further this thesis investigates shared control design for construction robots. Existing work has established that shared control can improve cycle times in nominal conditions. However, these methods can be too slow to relinquish control in off-nominal cases, when the operator needs to deviate from the nominally optimal trajectory due to unforeseen obstacles or other uncertainties. With an objective to incorporate such capability, this thesis proposes a new shared control technique that utilizes the operator's intent to quickly relinquish control in off-nominal conditions. Theoretical results with performance guarantees and improved obstacle reaction time are presented. Proposed design has been experimentally validated on Zermelo's navigation problem. The last part of the thesis introduces kernel observer for learning and inference of large-scale stochastic phenomena with both spatial and temporal (spatiotemporal) evolution. This work considers the problem of estimating the latent state of a spatiotemporally evolving continuous function using very few sensor measurements. The model consists of a dynamical systems prior over temporal evolution of weights of a kernel model. Theoretical results provide sufficient conditions on the number and spatial location of sensors required to guarantee state recovery. A lower bound on the minimum number of sensors required to robustly infer the hidden states is also derived. Finally, theoretical results for randomly selecting sensing or sampling locations based on the predictive kernel observer model are presented. Our approach outperforms state-of-the-art kernel based machine learning methods in numerical experiments on real world datasets.

Graduation Semester

2018-08

Type of Resource

text

Permalink

http://hdl.handle.net/2142/101651

Copyright and License Information

Copyright 2018 Harshal Maske

Owning Collections