Tutorial “Vision-Language-Action Models in Modern Robotics: Foundations and Research Frontier”

Abstract

This mini-course provides a coherent view of modern robotics through the lens of Vision-Language-Action (VLA) models. The first part covers how contemporary robotic systems are built and trained, from representations of observations and actions, data collection, and learning from demonstrations to simulation, real-world robots, and large-scale robotics datasets. We will discuss the typical VLA pipeline, major classes of robotic platforms and tasks, and the relationship between learning in simulation and in the real world. The second part moves from established methodology to the research frontier, including memory and partial observability, world models and future prediction, long-horizon planning, mobile and dexterous manipulation, adaptation, and generalization to new tasks, objects, and environments. The goal is both to explain how modern VLA systems work and to provide a map of the major open problems shaping current robotics research.