Event Location: GHC 6501

Abstract: The modular snake robots in Howie Choset’s lab provide an intriguing platform for research: they have already been shown to excel at a variety of locomotive tasks and have incredible potential for navigating complex terrains, but much of that potential remains untapped. Unfortunately, many techniques that are commonly used in robotics prove inapplicable to these snake robots because of either the robots’ complex, multi-modal locomotion dynamics – which are difficult to model – or their small size and frequent impacts, which preclude addition of many standard sensors. Therefore, I propose to extend research from other fields to address these challenges, simultaneously advancing the robotics literature as well as those fields from which the methods originate.

I am motivated to expand the capabilities of these robots after experiencing several failures and limitations in real world tests, caused by inadequate locomotive capabilities and situational awareness. To increase performance in these areas I propose the development of machine learning techniques as a means to build a suite of tools to make teleoperation simpler and more powerful. These tools include a library of gaits that have been optimized for use on various terrains, new non-gait motions to help the robot overcome obstacles, and methods for translating existing limited sensor data into a better description of the robot’s surroundings for the operator. This motivates solutions to the following more general theoretic problems:

1. Optimization of noisy, expensive, black-box functions, with extensions for multiple objectives, multiple fidelity models, and learning multiple related tasks.

2. Integration and generalization of demonstrated input along with automatic selection of useful basis primitives to reduce dimensionality of high-dimensional planning tasks.

3. Inference of useful environment parameters from noisy observations with an unknown observation model.

As an end goal, I hope to use these contributions to enable an operator to move a snake robot through a cluttered environment and locate a target, without knowing the layout of the environment or seeing the robot’s location.

Committee:Howie Choset, Chair

Jeff Schneider

Drew Bagnell

Jared Cohon

Stefan Schaal, University of Southern California