Episode Transformer: Model-based Episodic Reinforcement Learning
Keywords: Episodic RL, Model-based RL, Movement Primitives
TL;DR: We speed up episodic RL with movement primitives by learning a Transformer model to predict what will happen during an episode.
Abstract: Episodic Reinforcement Learning (ERL) with movement primitives (MPs) has recently achieved significant success, especially in sparse and non-Markovian reward scenarios. By reasoning directly at the trajectory level via MPs, ERL results in smoother, energy-efficient policies and improved exploration capabilities for many real-world tasks. However, these black-box optimization approaches have very poor data-efficiency making them impractical for real-world applications. To mitigate this drawback, we propose Episode Transformer, a model-based ERL algorithm. Here, we learn a transformer-based episodic world model. To perform control we train a policy, with trust region constraints, purely in the world model's imagination. We compare our approach to state-of-the-art step-based and episodic RL methods on a variety of challenging robotic tasks under dense, sparse, and non-Markovian reward settings. The results show that the Episode Transformer is able to learn high-quality policies that retain all the benefits of previous deep ERL methods while requiring up to 5x fewer environment samples.
Primary Area: reinforcement learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 9399
Loading