This paper presents a method for leveraging large-scale pre-trained models in the field of reinforcement learning (RL). Specifically, we propose InFOM (intention-conditioned flow occupancy models), probabilistic models that leverage flow matching to predict future states in RL environments where temporal dependencies are crucial. InFOM enhances the model's expressiveness by incorporating latent variables that capture user intent, enabling generalized policy improvement. We demonstrate that InFOM outperforms other pre-training methods on 36 state-based and 4 image-based benchmark tasks.
