Embodied AI (1/2)

Friday 5 Dec 2025, 14:15 → 15:45 Europe/Berlin

DO: JvF25/3-303 | BN: b-it/1.047

Introduction to Embodied Artificial Intelligence by Prof. Sven Behnke

Embodied Artificial Intelligence is an emerging paradigm that integrates perception, action, and reasoning within physical or simulated bodies, such as robots. In the lecture, we explore how embodiment enables agents to learn from interaction with their surroundings, adapt to dynamic environments, and ground abstract representations in sensorimotor experience. Core topics include embodied cognition, robot learning, multimodal perception, simulation-to-real transfer, and reinforcement learning for control. Embodied AI has numerous applications from autonomous driving to personal service robots.

 

Reinforcement Learning for Dynamic Robot Control by Julian Eßer

This lecture explores how reinforcement learning (RL) enables dynamic locomotion – balancing, agile maneuvers, and disturbance recovery – under contact-rich, uncertain conditions. We cover core design choices (observations, action spaces, rewards/curricula, safety) and practical sim-to-real strategies, including system identification, domain randomization, and hardware-in-the-loop adaptation. Recent advances that cut sample complexity and improve robustness are highlighted through the case study of learning dynamic control for a two-wheeled robot. Attendees will leave with actionable recipes, common pitfalls, and open resources to accelerate their own robot learning projects.