Methods for Teaching Diverse Robot Skills: Leveraging Priors, Geometry, and Dynamics
Loading...
Author(s)
Rana, Muhammad Asif
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Functioning in the real world requires robots to reason about and generate motions for execution of complex tasks, in potentially unstructured and dynamic environments. Early generations of robots were limited to simple tasks in controlled environments, where only a single skill was often required. To deal with the diversity of tasks and environments associated with the real world, robots should instead have access to a library of skills. Instead of pre-programming all the desired skills, a procedure which is cumbersome and often infeasible, it is beneficial to have a framework that allows robots to acquire new skills when required. One such framework is learning from demonstration, which provides a channel for robots to learn skills from everyday users. This dissertation provides methods for learning skills from human demonstrations.
Skill learning from human demonstrations carries certain challenges. Skills can be vastly different, enforcing a range of motion constraints. Human demonstrations are also often limited in number. Lastly, generalization of learned skills can be tied to generating motions that need to satisfy additional pre-specified constraints. These constraints can be associated with feasibility, requiring motions compliant with robot's kinematics and its environment, or they may be linked to coordination, requiring correlated motions of several robot body parts. To contend with the diversity of skills, the presence of feasibility and coordination constraints, and the scarcity of data, it is beneficial to impose structure in the skill representation. The structure incorporates domain knowledge in the representation, enabling desirable generalization even when access to large amounts data is hard.
The objective of this dissertation is to develop a family of techniques that allow robots to sample-efficiently learn diverse skills from human demonstrations, and subsequently generalize the skills to novel contexts while satisfying additional constraints that may exist, concerning the feasibility and coordination of robot motions. Each proposed method comes with a structured representation, suitable for tackling the challenges associated with a subset of skills. Specifically, we present: (i) a structured multi-coordinate cost learning framework coupled with an optimization routine, that generalizes skills requiring preservation of multiple geometric properties of motions, (ii) a structured prior representation employed in a probabilistic inference framework, geared towards generating optimal and feasibility-constrained motions, (iii) a stable dynamical system representation, suitable for learning skills aimed at motions that can react instantly to dynamic perturbation, and (iv) a tree-structured stable dynamical system which synthesizes multiple dynamical system into one, and learns skills dictating feasible and coordinated, yet reactive robot motions. As a preliminary to the aforementioned learning techniques, this dissertation also provides an over-arching benchmarking effort to identify the key challenges associated with skill learning from demonstration.
Sponsor
Date
2020-09-16
Extent
Resource Type
Text
Resource Subtype
Dissertation