Committee Members:

Professor Henrik Christensen (Advisor), ECE & UCSD

Stephen Balakirsky, PhD - GTRI

Professor Patricio Vela - ECE

Professor Aaron Ames - CalTech

Professor Dmitry Berenson - University of Michigan

Abstract:

This document introduces a novel technique for addressing the fragility of solutions that is common to configuration space-based planners. Both the state and the pre- and post-conditions of actions are represented as volumes of the configuration space. Feasibility of the various action types is determined by a subset test: if the set of possible states is a subset of the pre-condition set of a given action, then that action is feasible. This system is then incorporated into a planning system which is demonstrated on a number of household, scientific, and industrial application domains.

In the development of this framework, we build off of the idea of affordances: a high-level technique for reasoning about the space of latent action possibilities between an agent and its environment. In the language of affordances, our contributions are as follows:

* We formalize the computational notion of an affordance by introducing a set-based formulation of action pre-conditions and post-conditions.

* We define the feasibility of an action as an inclusion predicate between convex sets representing the possible system configuration and the parameterized action pre-conditions.

* We introduce a novel planning algorithm incorporating the previous contributions to jointly reason over information gathering and state transformation.

* We demonstrate the proposed system on a variety of simulated and real scenarios derived from household, scientific research, and industrial application domains.

This approach has a number of indirect benefits as well. First, the definition of action predicate sets depends on deep knowledge of the agent in question; as a result, the affordance representation specializes to a variety of agents while requiring only commonly-used mechanical and kinematic parameters. Second, as actions may serve to expand or contract the belief state, information-gathering actions may be planned without the need for e.g. a bespoke entropy minimization framework.

These benefits do incur some trade-offs, however. An explicit enumeration of possible actions is required, as are detailed models of their input-output behavior. These models may be developed from existing theory, via simulation, or by physical experimentation. In principle, such models could be discovered by self-experimentation or learning from demonstration, but these techniques are outside the scope of this work.