Fostering Intrinsic Motivation in Reinforcement Learning with Pretrained Foundation Models

In reinforcement learning, intrinsic motivation using pretrained foundation models, like CLIP, enhances exploration in sparse reward environments. Foundation models provide semantically rich embeddings for better state representation, improving sample efficiency and learning of optimal policies. Full state information boosts exploration, and CLIP embeddings often surpass those learned during training, accelerating the learning process. Experiments in the MiniGrid domain compare agents using different configurations of RIDE and FoMoRL, focusing on the impact of state information and episodic novelty terms on learning efficiency. Results highlight the effect of partial vs full observability on convergence. Future work aims to extend the approach to more environments and explore alternative methods leveraging foundation models.