LCSR Seminar: Physics-Grounded Artificial Intelligence

Description

Johannes Lachner, a postdoctoral researcher in the MIT-Novo Nordisk Artificial Intelligence Postdoctoral Fellows Program, will give a talk titled "Physics-Grounded AI: A Unifying Framework for Stable, Safe, and Human-Mimetic Robotic Physical Interaction" for the Laboratory for Computational Sensing + Robotics.

Lachner, who is also a lecturer and lab professor for one of MIT's flagship undergraduate mechanical engineering courses in instrumentation and measurement, works at the intersection of robotics, motor neuroscience, and AI. His research combines insights from human motor control to create robotic systems that support individuals with neurological and physical impairments.

Abstract:

Humans outperform robots in physical interaction, despite current advances in robot learning and control. AI-driven approaches are intrinsically unidirectional, constrained only by temporal causality and boundedness. Physical interaction, however, is dictated by reactive behavior governed by energy conservation, entropy, etc. Robot learning algorithms for physical interaction must therefore interface the information domain with the energy domain. In this talk, I present physics-grounded AI, a unifying framework that integrates coordinate-invariant, energy-consistent robot dynamics with machine learning and generative AI to enable safe, stable, and human-mimetic physical interaction. Inspired by human motor control, the framework enables modular robot control via motor primitives that form a compositional control language, allowing robots to learn interaction strategies from demonstration while preserving physical consistency. I will present my recent work on diffusion-based impedance learning, which bridges generative modeling with interactive control, enabling robots to adapt stiffness and damping autonomously, even across unseen tasks and hardware configurations. The focus application of this talk is robot-assisted physical rehabilitation, where my physics-grounded AI framework allows skilled physical and occupational therapists to imprint their treatment style onto robots, enabling personalized, stable, and safe assistance.