Artificial intelligence - driven mainly by large-scale statistical learning and increasingly rivaling human capabilities - is being incorporated into physical systems, tightly coupled with perception of the world and the ability to act on it. From intelligent infrastructure to highly autonomous robotic platforms, physical intelligence is spreading across our socio-economic fabric. Intelligent robots, including humanoids, are moving closer to our comfort zone and will soon shape outcomes in military operations, work environments, and homes. Ethics can no longer be an afterthought; it must be engineered as a core property of these systems. This talk presents a system-level perspective on ethics in physical intelligence, showing how ethical behavior can be engineered through perception, cognition, and action architectures, and through their interaction with human context and other forms of intelligence. It links ethical properties to system architecture across perception under uncertainty, decision-making in dynamic environments, action under physical irreversibility, and human-machine and machine-machine interaction, where trust, intent, transparency, and misinterpretation must be explicitly addressed. It examines learning and reasoning, and their implications for safety, data privacy, and ethical guarantees. It further addresses certifiable modes of operation, accountability in distributed system-of-systems architectures, distinctions between human-caused and system-generated ethical infringements, cultural variation in built-in ethics, and the need for human adaptation. Focusing on humanoid systems, it outlines principles for engineering, verification, and certification to help ensure that such systems operate safely, transparently, and ethically in the physical world.

Adrian grew up in Romania watching science-fiction films about space exploration and reading Asimov's books on humanoid robots, where he first encountered the laws of robotics. He later completed a PhD on humanoid learning by imitation and went on to spend nearly three decades at NASA's Jet Propulsion Laboratory, including about ten years leading the group that developed surface autonomy for Mars rovers. He is a NASA Institute for Advanced Concepts (NIAC) Fellow and a NASA-JPL Emeritus. He has also held leadership roles in IEEE, most recently serving as President of the IEEE Systems, Man, and Cybernetics Society. Adrian is passionate about humanoids both as a robotic workforce for building lunar industries and as systems that can assist people here on Earth.

In this talk, I demonstrate how discriminative models equipped with a softmax classifier can be reinterpreted as Energy-Based Models (EBMs). This perspective unlocks a range of discoveries across the machine learning spectrum, from adversarial robustness to Large Language Models (LLMs). First, I show how this framework sheds light on adversarial robustness, offering a deeper understanding of robust and catastrophic overfitting scenarios, as well as adversarial attacks. In the domain of LLMs, I will present a training-free method for detecting "hallucinations" by leveraging internal model values through an EBM lens. Finally, I discuss how this interpretation allows us to turn discriminative models into generative ones via model inversion. This enables the sampling of plausible data directly from discriminative models like robust classifiers, Vision-Language encoders like CLIP, and graphics systems whose encoding function is a differentiable renderer.

Iacopo Masi is an Associate Professor at Sapienza University of Rome and the founder of the OmnAI Lab. He previously held research faculty positions at the University of Southern California (USC) and the Information Sciences Institute (ISI). He has served as an Area Chair for major conferences, including CVPR, ICCV, and ECCV; as a General Chair for ICIAP 2025; and as a Program Chair for FG 2026. He is a recipient of the prestigious Rita Levi Montalcini Award in 2018. His research focuses on the intersection of computer vision and machine learning, with current interests in Robust and Trustworthy Machine Learning, AI Safety, and Generative AI.