AI That Moves, Adapts, and Learns: The Future of Embodied Intelligence

Sami Haddadin, a leading researcher in robotics and artificial intelligence, is shaping the future of AI—not the kind that exists only in code, but AI that engages with the real world. His work in human-centered robotics and embodied AI pushes beyond algorithms and datasets, building intelligent systems that don’t just process information but physically interact, adapt, and evolve.

As he prepares for his keynote at Columbia University’s AI Summit, Haddadin sat down with us to discuss why embodied AI is the next frontier—and why now is the moment to pay attention.

How is embodied AI different from traditional AI? 

Traditional AI, such as large language models, exists entirely in the digital realm—it processes data, recognizes patterns, and makes predictions. Embodied AI, by contrast, interacts with and reasons about the physical world.

“The key difference is that embodied AI learns through experience and interaction, much like humans,” Haddadin said. “It builds models of the world through sensory feedback and real-world interaction rather than just analyzing statistical data. One could say it is a curious discoverer in its own right.”

A robot equipped with embodied AI can do so much more than  just recognize objects—it autonomously learns to manipulate them, adapts to new conditions, and even refines its understanding over time. That’s what sets it apart.

“Because the world is constantly evolving, AI that exists in the physical world must account for this fluidity,” Haddadin explained. “There will never be a fixed ‘solution’ to embodied AI because its environment, like our own, is never fully explored and constantly changing.”

What excites you most about where the field is today? 

Haddadin believes embodied AI is at a pivotal moment.

“We are entering an era where AI is no longer confined to screens and datasets but will actively contribute to discoveries in the physical world,” he said. “Just as telescopes and microscopes expanded our ability to observe the universe, embodied AI could be the next great tool to extend human potential.”

One area he finds particularly thrilling is AI-assisted scientific discovery.

“Think of Iron Man’s AI assistant, Jarvis—not just a database, but an interactive partner in discovery,” Haddadin said. “Now imagine robots that collaborate with scientists, designing and testing hypotheses in real time, automating aspects of the scientific process.”

Embodied AI, he believes, will work alongside researchers to accelerate breakthroughs in medicine, material science, and even space exploration.

“By combining human ingenuity with machine adaptability, we could reach levels of problem-solving that neither could achieve alone.”

What are the biggest challenges in embodied AI?  

“The physical world is unpredictable,” Haddadin noted. Unlike digital environments, where AI models work with structured, controlled data, real-world interactions introduce sheer, unimaginable complexity. Bodies change, environments shift, and systems must continuously adapt.

“Just as a self-driving car must account for unexpected changes—pedestrians, weather, construction detours,” he said, “the same principle applies to robots interacting with humans in workplaces, hospitals, or homes.”

Embodied AI requires a fundamentally different approach—one that embraces uncertainty, continuous adaptation, and lifelong learning.

“It’s not a static problem to be solved,” Haddadin said. “It’s a dynamic process that will never be ‘finished’—just like human learning, or rather humanity’s collective learning.”

Will we ever have robots in every home?

For decades, science fiction has imagined a world where robots assist in homes and offices. Are we anywhere close to that reality?

“Many people imagine a future where robots are part of daily life,” Haddadin said. “But making that a reality is more challenging than it seems.”

While AI has reached superhuman levels in computational tasks—beating world champions in chess, solving complex protein structures—the physical world is another story.

“Robots struggle with basic tasks that humans take for granted,” he explained. “A child can pick up a toy from the floor without thinking, but for a robot, that involves an incredible level of complexity—perception, movement planning, adapting to the object’s shape and weight.”

Still, progress is happening. “Robots are becoming better at manipulating objects, learning from their environments, and functioning in real-world conditions,” Haddadin said. “The challenge isn’t just technical—it’s about safety, cost, and adaptability. We may not have fully autonomous household robots yet, but we are getting closer every day.”

What is the core message of your keynote at the Columbia AI Summit? 

Haddadin paused before answering.

“That’s quite a question,” he said. “But if I had to sum it up, it’s really about what it means to develop a truly artificial intelligence in the physical and embodied realm. How do we build AI systems that can connect to different embodiments and solve real-world physical problems?”

Beyond that, Haddadin wants to push the conversation further.

“If we can start answering that question—at least making progress on it—the hypothesis is that we can take steps toward something much bigger: the automation of science itself. That’s what I want to explore in the talk.”

Final Thoughts

Embodied AI is not just another iteration of machine learning—it’s a fundamental shift in how AI interacts with the world. Haddadin’s work challenges us to rethink intelligence, adaptability, and what machines can truly accomplish.

Don’t miss his keynote on March 4, 2025, at the Columbia AI Summit.