Recent Progress on Foundation Model Supervision for Robot Learning

Watch a 53-minute seminar presentation from MIT where PhD student Jason Ma discusses innovative approaches to robot learning using foundation models. Explore how foundation models from non-robot domains can provide scalable supervision for diverse robotic tasks without extensive training data or hand-engineered controllers. Learn about Eurekaverse, a Large Language Model-based environment curriculum generation algorithm that enables robots to acquire complex parkour skills in real-world settings. Discover Generative Value Learning, a novel approach utilizing visual language models for universal value function learning. Gain insights into cutting-edge research that has earned recognition as a Best Paper Finalist at ICRA 2024 and been named among NVIDIA's Top 10 Research Projects of 2023, while understanding how these advances are bringing us closer to achieving general-purpose robotics.