Research Overview

Written on October 12th, 2021 by Geoffrey Clark

My research focuses primarily on adding Bayesian filtering and optimal control concepts to probabilistic machine learning paradigms.

In short, classical filtering and optimal control methods are typically slow and difficult to tune for systems which are both high-dimensional and nonlinear (such as human-robot interaction) while machine learning methods do not provide guarantees for stability and robustness in learned controllers. By merging ideas from both I am working to create intelligent robot control methods which are an order of magnitude faster than conventional optimal controllers.

One excellent application of my research is in the control of powered prosthetics. For example I enabled fast and accurate prostheses control by training a model to predict high-fidelity control signals from low-fidelity sensors such as IMUs (ICRA 2020). Going a step further I incorporated model predictive control to ensure stability and robustness while continuously looking ahead to possible future states and modifying the robots interaction to insure the amputee walks in ways which take less energy, increase stability, or limit the risk of osteoarthritis (CoRL2020). I am now working towards a unified motor primitive approach that tackles both optimal perception and optimal action.

Autonimous robotics is another interisting aplication focus for my work on pairing optimal perceptions with optimal actions. Without humans in the loop a robot – whether to play chess or an autonimous legged system – must still evaluate the environment and formulate a model of how to interact with it. This interaction, environment to robot and robot to envronment forms a loop that allows for optimization of both sides by the robot.