Seminar - Bayes' Rule Gives You Wings: Probabilistic Autonomy for Boundedly Rational Robots - Oct. 23
Nisar Ahmed
Associate Professor, Smead Aerospace
Friday, Oct. 23 | 10:40 A.M. | AERO 111
Abstract: Autonomous robotic systems are appearing across land, sea, air, and space. To perform sophisticated tasks in these domains, robots face the common problem of intelligently reasoning about and acting against uncertainties. However, they must also deal with constraints on computing, data, size, power, communications, and safety. As boundedly rational agents, autonomous robots must therefore also be capable of examining and augmenting their limited abilities, if they are to realistically satisfy expectations of self-directedness, self-sufficiency, interactivity, and reliability in achieving desired goals.Ìý
I will present developments in probabilistic modeling and algorithm design for realizing these capabilities, led by my research group at the Cooperative Human-Robot Intelligence (COHRINT) Lab. After a brief overview of our work on semantic collaborative active sensing and competency self-assessment for human-robot teams, I will focus on Bayesian state estimation methods for decentralized data fusion (DDF) problems in heterogeneous multi-robot teams. These problems ask how robots can soundly combine local probabilistic beliefs about shared task variables in a peer-to-peer manner, despite having different local state space models and asymmetric sensing, computing, and communication abilities. I will describe our factor graph DDF (FG-DDF) framework and show how it offers the most promising general solution to such problems. FG-DDF exploits novel insights about conditional independence properties of local robot beliefs to avoid information double-counting and drastically reduces computation and communication costs for decentralized data fusion. This permits realistic modular multi-robot decentralized state estimation solutions for large problem domains, such that required costs scale with the number of shared variables (not number of robots). Results from simulation and hardware studies involving GPS-limited multi-robot cooperative navigation and multi-object tracking problems will be presented, driven by applications to search and rescue, defense monitoring, and space exploration.Ìý
Bio: Nisar Ahmed is an Associate Professor in the Â鶹Ãâ·Ñ°æÏÂÔØAerospace Engineering Sciences Department, and is Director of RECUV and the COHRINT Lab. He received his B.S.E. from Cooper Union in 2006 and his Ph.D. in Mechanical Engineering from Cornell in 2012, where he was also a postdoctoral researcher from 2012 to 2014. He was awarded the 2011 AIAA GNC Conference Best Paper Award; ASEE Air Force Summer Faculty Fellowship in 2014; 2018 Aerospace Control and Guidance Systems Committee (ACGSC) Dave Ward Memorial Lecture Award; and Smead Aero Faculty Fellowship (2021-2023) and Outstanding Graduate Teaching and Mentorship Award (2021). His work has been supported by the Army, Air Force, DARPA, Navy, NASA, Space Force, and multiple industry sponsors. He is a Member of IEEE and the AIAA Intelligent Systems Technical Committee, and Â鶹Ãâ·Ñ°æÏÂÔØSite Director of the NSF Center for Autonomous Aerial Mobility and Sensing (CAAMS).
Ìý