TBD

Abstract: Theory of Mind Modeling in Search and Rescue Teams

Theory of Mind (ToM) refers to the ability to make inferences about other’s mental states. Such ability is fundamental for human social activities such as empathy, teamwork, and communication. As intelligent agents come to be involved in diverse human-agent teams, they will also be expected to be socially intelligent in order to become effective teammates. In this talk, we describe motivations, challenges and advantages of ToM for single human and human teams. We present a computational ToM model which observes team behaviors and infers their mental states in an urban search and rescue (US&R) task. Our modular ToM model approximates human inference by explicitly representing beliefs, belief updates, and action prediction/generation. We present experimental results that compare the model’s performance with human observers that are asked to make the same inferences. We also discuss the complementary problem of Human Theory of Machines as well as open research problems.

Abstract: Symbolic Mental Models as a Lingua Franca for Supporting Explainable & Advisable Human-AI Teaming

Despite the surprising power of many modern AI systems that often learn their own representations, there is significant discontent about their inscrutability and the attendant problems in their ability to interact with humans. While alternatives such as neuro-symbolic approaches have been proposed, there is a lack of consensus on what they are about. There are often two independent motivations (i) symbols as a lingua franca for human-AI interaction and (ii) symbols as system-produced abstractions used by the AI system in its internal reasoning. The jury is still out on whether AI systems will need to use symbols in their internal reasoning to achieve general intelligence capabilities. Whatever the answer there is, the need for (human-understandable) symbols in human-AI interaction and teaming seems quite compelling. In particular, humans would be interested in providing explicit (symbolic) knowledge and advice -- and expect machine explanations in kind. This alone requires AI systems to maintain symbolic mental models for supporting explainable and advisable teaming with them. In this talk I will outline my group's research efforts in realizing this vision.

Short Bio:

Subbarao Kambhampati is a professor of computer science at Arizona State University. Kambhampati studies fundamental problems in planning and decision making, motivated in particular by the challenges of human-aware AI systems. He is a fellow of Association for the Advancement of Artificial Intelligence, American Association for the Advancement of Science, and Association for Computing machinery, and was an NSF Young Investigator. He served as the president of the Association for the Advancement of Artificial Intelligence, a trustee of the International Joint Conference on Artificial Intelligence, the chair of AAAS Section T (Information, Communication and Computation), and a founding board member of Partnership on AI. Kambhampati’s research as well as his views on the progress and societal impacts of AI have been featured in multiple national and international media outlets. He can be followed on Twitter @rao2z.

Abstract: How can we learn human models more efficiently?

In this talk, I will discuss two of our recent works on enabling robots to more effectively learn latent properties of humans. The first work is on human-robot assistive communication by planning over a learnt human model. Here, the key problem is that human models are difficult to obtain. We take inspiration from social projection theory and show that learning differences from the robot’s self model is more data-efficient than learning an entire human model from scratch. This leads to models that are effective for planning communication, as we will demonstrate in experiments. The second work is inverse reinforcement learning (IRL), which is important for value alignment and learning from demonstration. A fundamental issue with IRL is unidentifiability: there are multiple reward functions that can give rise to a given policy or observed behavior. Here, we propose a Bayesian Optimization (BO) approach to IRL that efficiently explores the reward function space to identify multiple solutions consistent with expert demonstrations. However, we find that a naive direct use of BO leads to unsatisfactory results. Our key insight is to leverage policy invariance to derive a new projection kernel, which leads to significantly better results in experiments.

Short Bio:

Harold Soh is an Assistant Professor in the Department of Computer Science at the National University of Singapore (NUS), where he directs the Collaborative Learning and Adaptive Robots (CLeAR) lab. Harold completed his Ph.D. at Imperial College London on online learning for assistive robots. Harold’s current research focuses on machine learning and decision-making for trustworthy collaborative robots. His work, which spans cognitive modeling (human trust) to physical systems (tactile perception with novel e-skins), has been recognized with a best paper award at IROS’21 and nominations at RSS’18, HRI’18, RecSys’18, and IROS’12.

Short Bio:

Dr. Cangelosi has a background in both cognitive science and artificial intelligence and has made significant contributions to Human-Robot Interaction by introducing hybrid cognitive models that capture how empathetic trust is built towards an agent. His research areas are focused on language grounding and embodiment in humanoid robots, developmental robotics, human-robot interaction, and on the application of neuromorphic systems for robot learning. His latest book ‘Cognitive Robotics’ (MIT Press), was co-edited with Minoru Asada and published open access in 2022.