Hugging Face Integrates the Decision Transformer Library

Hugging Face has integrated Decision Transformer, an Offline RL methodology, into the transformers library and the Hugging Face Hub.

Decision Transformer redefines reinforcement learning as a conditional-sequence modeling problem. While traditional reinforcement learning learns a value function to maximize reward, this model uses an autoregressive approach that takes the desired return, past states, and actions as input to generate future actions to achieve the target return.

This approach shifts the reinforcement learning paradigm toward generative trajectory modeling. This enables offline reinforcement learning, allowing an agent to learn using only existing datasets (human demonstrations or records from other agents) without directly interacting with the environment.