Peng Yu @ AeroAstro, MIT

I am a second year PhD student working with Professor Brian Williams at the MIT Computer Science and Artificial Intelligence Lab. My research focuses on the collaboration between humans and autonomous systems in planning and execution tasks, especially under over-subscribed and uncertain situations. As the complexity of modern planning tasks increases, we are increasingly relying on autonomous systems to generate plans for our requirements and uncertainty in the environments. Often, the systems are able to detect and signal when the planning problem is over-subscribed, but not providing resolutions to recover from the failure. The user needs to find the cause of failures and adjust the requirements to resolve them. This has motivated me to investigate a collaborative approach that enables the autonomous systems to provide guidance for the users to cope with the over-subcription and uncertainty, as oppose to just noticing them. When a plan becomes over-subscribed due to hardware failure or increasing uncertainty in the environment, the autonomous system will find the causes and generate resolutions, and explain them to the operators to help them restore the feasibility of the plan.


There are two major research questions in my approach: (a) automatic diagnosis of over-subscribed temporal plans, and (b) efficient communication of resolutions and decisions between human and autonomous systems. I have mainly focused on (a) and developed a relaxation algorithm for over-subscribed plans with temporal uncertainty. It uses a conflict-directed approach and can enumerate continuous relaxations in preferred orders. When there are activities with uncontrollable durations, the algorithm can tighten their bounds to restore controllability. I am currently extending the algorithm to deal with chance-constrained problems with probabilistic durations and chance constraints. This allows autonomous systems to resolve over-subscribed problems through negotiating acceptable risk levels with the user, and makes the algorithm applicable to more real-world scenarios.


I received Bachelor of Engineering degree in Mechanical Engineering from the Hong Kong University of Science and Technology in June 2010. For my final year design project, four teammates and I designed a mini wind tunnel with control software packages to evaluate various airfoil designs in low speed air flow. I also worked on several robotic projects during my undergraduate study, include the ABU Asia Pacific Robot Contest and Personal Robot 2. For more details, see my CV. You may find me through email (yupeng AT mit.edu) or in the MERS lab (32-224, MIT).


Figures on top: The left one is a Personal Air Vehicle in simulation software X-Plane 9. The vehicle is controlled by a temporal plan executive, Kirk, and my collaborative diagnosis algorithm. The right one is the TERRAFUGIA flying car. I wish I could have my algorithm implemented on the car and make it a truly automatic air taxi.