Nicklas Hansen - World Models on an Academic Budget

Recent progress in AI can largely be attributed to the emergence of large models trained on large datasets. However, teaching AI agents to reliably interact with our physical world has proven challenging and, consequently, this new paradigm has not yet materialized as much in robotics as in related areas. World models provide a promising framework for modeling physical interaction and are quickly gaining traction in industry. However, how to build a good world model is still mostly an open question. In this talk, I will share my perspective on the future of world models, how academia can get involved, as well as our current steps toward more capable models. Concretely, I will discuss our work on TD-MPC, a highly data-driven approach to world models that scales with data and model size, improves autonomously through interaction, and can be trained on a modest budget. I will cover its algorithmic foundations and application to diverse decision-making problems in embodied AI, and conclude the talk with a discussion of open problems. Nicklas Hansen is a PhD candidate at University of California San Diego advised by Professors Xiaolong Wang and Hao Su. Their research focuses on developing generalist AI agents that learn from physical interaction. Nick has spent time at NVIDIA Research, Meta AI (FAIR), as well as Berkeley AI Research, and received their BS and MS degrees from Technical University of Denmark. They were a recipient of the 2024 NVIDIA Graduate Fellowship, and their work has been featured at top venues in machine learning and robotics. This session is brought to you by the Cohere Labs Open Science Community - a space where ML researchers, engineers, linguists, social scientists, and lifelong learners connect and collaborate with each other. We'd like to extend a special thank you to Rahul Narava Gusti Winata, Leads of our Reinforcement Learning group for their dedication in organizing this event. If you’re interested in sharing your work, we welcome you