IROS 2025 Keynotes - Embodied Intellgence: Fumin Zhang

Why It Matters

By integrating generative AI with proven search theory, robots can conduct faster, explainable searches, dramatically lowering the time and expense of critical rescue and inspection missions.

Key Takeaways

• Traditional search theory relies on only two strategies.
• Unified framework merges Bayesian and source‑seeking using model predictive control.
• Generative AI enables camera‑based search via 3D Gaussian splatting.
• Bayesian optimization creates surrogate similarity functions for real‑time searching.
• Explainable control outperforms black‑box reinforcement learning in robotic searches.

Summary

The IROS 2025 keynote by Fumin Zhang examined how robots can perform high‑stakes search and rescue tasks by marrying classic search theory with modern generative AI and control techniques. Zhang highlighted that, despite a half‑century of research, the field has traditionally relied on just two paradigms: Bayesian (Bashian) probability‑map search and source‑seeking based on signal gradients. He argued that the rapid advances in AI‑driven perception now allow these abstract models to be applied to visual sensors, opening a path to fully autonomous, explainable search behaviors.

Zhang dissected the core components of search theory—sensor models, prior probability distributions, and detection‑gain functions—and showed how they can be cast as a reward in reinforcement‑learning or, more transparently, as a cost in model‑predictive control (MPC). By updating the prior with Bayesian learning and integrating detection gain, his unified framework simultaneously executes structured (Bayesian) and gradient‑climbing (source‑seeking) actions. This synthesis was demonstrated on a mobile robot locating a coffee machine and on a custom‑built blimp hunting a miniature dog model, illustrating both the theoretical soundness and practical feasibility.

Key experimental insights included the use of 3D Gaussian splatting to synthesize future camera views, enabling a “cookie‑cutter” detection model for visual similarity. Initial trials using raw similarity scores were accurate but painfully slow (15× slower than real time). By treating the similarity field as a rate function and approximating it with Bayesian optimization, computation dropped to roughly 30 seconds, making real‑time deployment viable. The approach also revealed why animal‑inspired random searches sacrifice efficiency for reduced localization demands, a trade‑off robots can now navigate deliberately.

The broader implication is a shift toward explainable, data‑efficient robotic search that leverages AI without abandoning the rigor of decades‑old theory. Practitioners can replace costly human‑led sweeps in maritime, disaster, and industrial settings with robots that predict their visual observations, adaptively refine target priors, and provide transparent decision logs—addressing safety, cost, and accountability concerns that have limited wider adoption.

Original Description

"Keynote Title: ""Bayesian Learning and Bio-Inspired Autonomous Search""

Speaker Biography

Dr.Fumin ZHANG is Chair Professor and Director of the Cheng Kar-Shun Robotics Institute at the Hong Kong University of Science and Technology. He was the Dean’s Professor in the School of Electrical and Computer Engineering at the Georgia Institute of Technology during 2007-2023. He received a PhD degree in 2004 from the University of Maryland (College Park) in Electrical Engineering and held a postdoctoral position in Princeton University from 2004 to 2007. His research interests include mobile sensor networks, maritime robotics, control systems, and theoretical foundations for cyber-physical systems. He received the NSF CAREER Award in September 2009 and the ONR Young Investigator Program Award in April 2010. He is currently serving as the co-chair for the IEEE RAS Technical Committee on Marine Robotics, associate editors for IEEE Transactions on Automatic Control, and IEEE Transactions on Control of Networked Systems, IEEE Journal of Oceanic Engineering, and International Journal of Robotics Research. He is Fellow of IEEE.

Abstract

The employment of robotic platforms and autonomy might significantly increase the efficiency and reduce the risk to humans in search and rescue missions. Motivated by insights from the autonomous collective foraging behaviors performed by animals in aquatic environments, this talk introduces models and provable strategies from control theory and robotics towards bio-inspired autonomous search operations. The bio-inspired methods generalize to a Bayesian learning framework where insights from biology are well justified by systems theory such as reachability, consistency, and optimality. Experimental effort with promising results demonstrates that bio-inspired autonomy might be preferred in aquatic environment that features severe limitation in communication, localization, and power consumption."