Worlds as simulators
Can a neural network learn an environment well enough to replace it? We study action-conditioned video and world prediction, neural game and robotics simulators, and controllable video generation.
Generative models have learned to produce strikingly realistic images and video. The next step is worlds: models that capture how environments evolve, respond to actions, and stay coherent over time. GLOW (Generative Learning of Worlds) Lab, led by (Incoming) Prof. Seung Wook Kim, studies these visual world models in both of their roles — as simulators that can stand in for reality, and as policies that let agents plan and act — along with the representations that make both possible. World models like these are essential for physical AI: any system that must sense, predict, and act in the real world needs one.
Can a neural network learn an environment well enough to replace it? We study action-conditioned video and world prediction, neural game and robotics simulators, and controllable video generation.
A model that predicts the world can also decide in it. We use world models as the backbone of policy models: planning with learned dynamics, video-as-policy, and grounding control in inverse dynamics.
What makes a latent space good for prediction and control? We investigate the representations world models learn and are built on — video VAEs and latent-space design, predictive representation learning, and latents that are rich in dynamics.
Wonjun Chang, Sanghyeop Kim, Kyaw Ye Thu, Minseo Kim, and Yeonwoo Shin join the lab as research interns — welcome!
GLOW Lab opens at the Kim Jaechul Graduate School of AI, KAIST.
* indicates equal contribution