Talk Topic:
Distributed AI Agents for Scientific Discovery and Real-World Decision-Making
Speaker Bio:
Saptarashmi Bandyopadhyay is a Tenure-Track Assistant Professor of Computer Science at the City University of New York at the City College of New York and the Graduate Center. He graduated with his Ph.D. in Computer Science at the University of Maryland, College Park (UMD) advised by Prof. John Dickerson and Prof. Tom Goldstein, in Summer 2025. His research on Multi-Agent AI for Autonomous Decision Making in the Real World addresses the challenges and opportunities of building AI Agents to plan, reason, and navigate in AR/VR, Supply Chains, Recommender Systems, Robotics, Self-Driving Cars, Climate Conservation, and other domains. He works with Reinforcement Learning, Imitation Learning, Model Predictive Control, LLMs, VLMs, and Game Theory algorithms to train AI Agents with Social Intelligence to take actions and provide insights at scale. He has been a Ph.D. Student Researcher at Google Augmented Reality and Google DeepMind in the Multimodal Conversational AI and Astra AR teams creating Multimodal (Audio, Vision and Language) AI Agents to proactively assist users. At UMD, he has been the Lead PhD RA of a DoD project on Explainable AI Agents. Saptarashmi has published twenty-six research papers in top AI venues including AAAI, ACM AAMAS, NeurIPS, EMNLP, ACL, SPIE, and others. He chaired the Multi-Agent AI in the Real World Workshop at AAAI-25 and created the MARL Seminar at UMD, hosting prominent speakers from industry and academia including Turing Award Laureates. Previously, Saptarashmi was an AI Resident at Google X, and did research internships at CNRS LORIA and INRIA in France (as a Charpak Scholar), and DFKI and the University of Saarland in Germany. He is a Do-Good Fellow and Dean’s Summer Fellow at UMD. Further research details can be found on his websites https://sites.google.com/view/saptarashmi/about and https://www.gc.cuny.edu/people/saptarashmi-bandyopadhyay
Talk Abstract:
Artificial Intelligence (AI) Agents are increasingly being deployed in Robotics, Augmented Reality/Virtual Reality, Self-Driving Cars, Scientific Discovery, Network Communications, and other domains. Agents need to reliably cooperate with humans using algorithms such as Multi-Agent Reinforcement Learning (MARL) and Imitation Learning (IL). In this talk, Saptarashmi will introduce an imitate-then-commit algorithm for AI Agents by unifying concepts from IL and Computational Game Theory to cooperate and align in settings where they have similar goals but different priorities. Guarantees on this approach are stronger than a naive reduction of the alignment problem to IL. Saptarashmi will then share a Multimodal Agentic Model Predictive Control framework to allow fine-grained tuning of Imitation Learning demonstrations, using VLMs, to train autonomous vehicles with better spatio-temporal reasoning and improved control dynamics. Next, Saptarashmi will share real-world applications of AI Agents, including YETI (YET-to-Intervene) Multimodal Agents which efficiently detect when to autonomously intervene while interacting with users in AR for planning, guidance, navigation, fixing mistakes or other tasks. He will introduce his research on improving automated scientific discovery in protein structures, neuroscientific modeling, accelerated photonics and material design. His focus on improving on scalable and dynamic exploration and dynamic prediction of protein structures has led to domain-specific improvements over Alphafold. He will introduce solutions to the problem of energy and system efficiency for these intelligent agents with a Multi-Agent AI Designer to reliably assist and stabilize these challenges. He will also share his research on AI Agents for Climate Conservation, Education, Supply Chain Orchestration, Building Engineering Automation and other areas. Saptarashmi will highlight the importance of training such AI Agents at scale efficiently and introduce JAXMARL, the fastest open-source MARL library with up to 12,500× speedup over alternatives. Together, his talk shows the promise of efficient and generalizable Deep Learning algorithms, guiding AI Agents and Multi-Agent Decision Making with human interaction to solve real-world problems which add new capabilities to AI Agents such as planning, reasoning and navigation while optimizing system performance by distributed processing.