[Robot Learning Team Seminar] Guest Talks by Prof. Giovanni Beltrame and Prof. Glen Berseth (Mila)

15:30-16:15
Speaker: Prof. Giovanni Beltrame
Full Professor at Polytechnique Montréal & Mila Member

Talk Title:
Physical AI for Field Robotics: Challenges and Applications

Abstract:
Physical AI promises machines that perceive, reason, and act in the unstructured world: search-and-rescue sites, agricultural fields, underground mines, and disaster zones where conditions are unpredictable and infrastructure is absent. Yet bridging the gap between learned models and reliable physical behavior remains one of the central challenges in modern robotics.

This presentation examines how recent advances in physical AI can be brought to bear on field robotics, where robots must operate autonomously in environments that resist the clean assumptions of the lab. Drawing on lessons from both single-robot and multi-robot systems, it explores how perception, coordination, and learning come together—or fail to—when deployed outside controlled settings.

Making physical AI practical in the field requires confronting several interconnected problems. Robots must perceive and map their surroundings accurately enough to navigate safely, often without GPS or prior maps. They must communicate and coordinate reliably in environments with no existing infrastructure, such as underground caves or remote terrain. They must manage limited resources—battery, computation, and bandwidth—while operating for extended periods.

A particular focus of this talk is the challenge of visual models. Modern vision systems, including large pretrained and foundation models, offer remarkable capabilities but can degrade in unpredictable ways under distribution shift, adversarial lighting, motion blur, and the long tail of conditions encountered in field deployment. Understanding when these models can be trusted—and how to build systems that remain robust when they fail—is essential for safe autonomy.

The research presented here offers concrete approaches to these challenges, validated through field missions in some of the most demanding environments on Earth. The long-term goal is a future where autonomous robots, endowed with physical AI, serve as reliable partners in exploration, monitoring, and intervention across domains where they are needed most.

BIO:

Prof. Giovanni Beltrame is a Full Professor at Polytechnique Montréal, a Mila member, and Director of the MIST Lab. His expertise spans AI-driven robotics, swarm intelligence, and space technologies, with a strong track record of collaboration with NASA, ESA, and the Canadian Space Agency on autonomous systems operating in challenging environments.

16:30-17:00
Speaker: Prof. Glen Berseth
Associate Professor, Université de Montréal and Mila Quebec AI Institute

Talk Title:
From Trial and Error to Foresight: Building Robots That Learn, Generalize, and Act

Abstract:
How do we build robots and AI agents that can learn to act intelligently in the real world? In this talk, I'll share several recent projects from my lab that chip away at this problem from different directions. I'll start with a basic question about how machines learn from trial and error: when a learning algorithm struggles, is it because it isn't trying enough new things, or because it can't properly use the experience it already has? Getting this right changes how we design better learning systems. From there, I'll show how giving robots a more structured way of "seeing" the world—by recognizing objects and their boundaries—helps them learn visual tasks faster and more reliably, and how teaching a robot's internal representations to recognize patterns that recur across different tasks helps it generalize to situations it hasn't seen before. I'll then turn to today's large, pre-trained robot control models—systems trained on huge amounts of demonstration data—and show two ways to make them more capable after training: by having them "think ahead" and search over possible action sequences before acting, and by relabeling existing robot demonstration data so that models become far more robust when the task changes slightly. Finally, I'll describe a way to automatically generate the reward signals that guide a robot's learning using large language models, removing much of the tedious hand-engineering that has traditionally been required. Together, these projects reflect a common thread: making robots and AI agents that learn faster, generalize better, and require less manual effort to teach—bringing us closer to machines that can reliably help us in everyday, real-world settings.

BIO:
Glen Berseth is an associate professor at the Université de Montréal, a core academic member of the Mila - Quebec AI Institute, Canada CIFAR AI chair, and co-director of the Robotics and Embodied AI Lab (REAL). He was a Postdoctoral Researcher at Berkeley Artificial Intelligence Research (BAIR), working with Sergey Levine. His current research focuses on machine learning and solving real-world sequential decision-making problems (planning/RL), such as robotics, scientific discovery and adaptive clean technology. The specifics of his research have covered the areas of human-robot collaboration, generalization, reinforcement learning, continual learning, meta-learning, multi-agent learning, and hierarchical learning. Dr. Berseth has published across the top venues in robotics, machine learning, and computer animation in his work. He also created a new course on foundational models and scaling reinforcement learning for robotics at Université de Montréal and Mila, covering the most recent research on machine learning techniques for creating generalist agents. He has also co-created a new conference for reinforcement learning research.