By adding explicit world modeling to the critic function, WCM helps robot learning systems better understand how observations change over time, leading to more reliable value estimates and improved performance on both seen and unseen tasks.
This paper introduces World Critic Model (WCM), a new approach for training robot control systems that combines vision, language, and action learning. The key innovation is having the critic (value estimator) explicitly learn to predict future states alongside estimating action values, rather than just predicting scalar returns.