Skip to main content

Physical Information & Active Learning

How does a robot use its physical capabilities to actively explore, sense, and learn about itself and its environment? This research is about developing control policies to exploit the physical motion of robotic systems for active learning and sensing. We are interested in understanding how information can be used to drive decision-making and how one can generate control actions with respect to principles of information theory and measures of information.

Autonomous Control Synthesis based on Information and Statistical Distributions

This research is focused on understanding how information and statistics can be used to drive decisions and how one can generate control actions with respect to principles of information theory and statistics. Our current research focus is on understanding how these principles can be incorporated into the objectives for motion tasks and deriving real-time algorithms that can robustly generate efficient trajectories that still obey the system constraints and dynamics. For a quick, high-level overview of our approach, watch our video "Autonomous Robot Drawing: From Distributions to Actions Using Feedback"

Autonomous Robot Drawing

This project is focused on generating methods to translate information via controls for physical motion. By using information density distributions to synthesize controls, we can enable robots to autonomously render an image based on the visual information. The full video of all the robot drawings can be found below. 

Sawyer Robot drawing a portrait of Abraham Lincoln

For more information, read: 

https://arxiv.org/abs/1709.02758

Watch the NSF "Science Nation" feature video of this work here.

Autonomous Shape Estimation

This project focuses on developing an algorithm for autonomous shape exploration using sensory contact information. In particular, we are developing algorithms for real-time shape exploration that does not rely on prior knowledge of shapes or the number of objects in the workspace, but can adapt control actions in real-time as information is obtained.

Shape estimation of two different objects over time using a binary contact sensor and ergodic exploration

For more information, read:

https://arxiv.org/abs/1709.01560

Ergodic Control for Exploration

This research is focused on using principles from statistics to understand how biological systems plan movements with respect to the movement goal and how this understanding can be applied to robotics and rehabilitation strategies. In particular, we are interested in how the objective for motion tasks, for example gathering sensory information, drives the way in which a system plans movement.  This research relies on the use of a mathematical concept from the field of ergodic theory, ergodicity, which provides a quantitative way of understanding the spatial statistics of a motion. 

For more information, read:

https://arxiv.org/abs/1708.09352

Trajectory Optimization for Parameter Estimation

This project focuses on developing algorithms for efficiently estimating parameters of a system. By incorporating information metrics into the objective function, we are able to actively identify model parameters of a system and adapt the model and controls in order to accomplish a task. We are interested in investigating efficient ways to generate control actions in real-time and how metrics such as the Fisher information metric can be used to efficiently drive decisions to facilitate active parameter estimation.

For more information, read:

https://arxiv.org/abs/1709.03474

https://arxiv.org/abs/1709.03426

People

Todd Murphey
Ian Abraham
Ahalya Prabhakar
Andrew Wilson

Related publications

Fitzsimons, K., A. Maria Acosta, J. Dewald, and T. D. Murphey, "Ergodicity reveals assistance and learning from physical human-robot interaction", Science: Robotics, vol. 4, issue 29, 04/2019. Google Scholar

Abraham, I., and T. D. Murphey, "Decentralized Ergodic Control: Distribution-Driven Sensing and Exploration for Multiagent Systems", IEEE Robotics and Automation Letters, vol. 3, pp. 2987-2994, Oct, 2018. Google Scholar

Abraham, I., A. Prabhakar, and T. D. Murphey, "Active Area Coverage from Equilibrium", Workshop on the Algorithmic Foundations of Robotics (WAFR), 2018. Google Scholar 

Mavrommati, A., E. Tzorakoleftherakis, I. Abraham, and T. D. Murphey, "Real-Time Area Coverage and Target Localization Using Receding-Horizon Ergodic Exploration", IEEE Transactions on Robotics, issue 34, pp. 62-80, Jan-01-2017, 2018. Google Scholar 

Abraham, I., A. Prabhakar, M. J. Z. Hartmann, and T. D. Murphey, "Ergodic Exploration using Binary Sensing for Non-Parametric Shape Estimation", IEEE Robotics and Automation Letters, vol. 2, issue 2, pp. 827-834, 2017. Google Scholar

Miller, L. M., Y. Silverman, M. A. MacIver, and T. D. Murphey, "Ergodic Exploration of Distributed Information", Transactions on Robotics, vol. 32, issue 1, pp. 36-52, 2016. Google Scholar 

Prabhakar, A., A. Mavrommati, J. A. Schultz, and T. D. Murphey, "Autonomous Visual Rendering using Physical Motion", Workshop on the Algorithmic Foundations in Robotics (WAFR) 2016, 2016. Google Scholar

Wilson, A. D., J. A. Schultz, and T. D. Murphey, "Trajectory Optimization for Well-Conditioned Parameter Estimation", IEEE Transactions on Automation Science and Engineering, vol. 12, issue 1, pp. 28-36, 2015. Google Scholar 

Miller, L. M., and T. D. Murphey, "Trajectory Optimization for Continuous Ergodic Exploration", American Control Conference (ACC), 2013. Google Scholar

Wilson, A. D., and T. D. Murphey, "Optimal Trajectory Design for Well-Conditioned Parameter Estimation", IEEE Conference on Automation Science and Engineering (CASE), pp. 13-19, 2013. Google Scholar 

Miller, L. M., and T. D. Murphey, "Optimal contact decisions for ergodic exploration", IEEE Int. Conf. on Decision and Control (CDC), 2012. Google Scholar

Back to top