Hugging Face adds Decision Transformer to transformers with pretrained Gym checkpoints

AI Impact Summary

Hugging Face has integrated Decision Transformer, an offline reinforcement learning model, into the transformers library and the Hub. It ships nine pretrained checkpoints for continuous-control tasks (e.g., Hopper, Walker2D, HalfCheetah) and provides example code and a DecisionTransformerModel interface to load from the hub. Practitioners can condition on return-to-go, past states, and actions to generate future actions, enabling offline RL experimentation without live Environment interaction. Teams should align their data pipelines to the required inputs (state_dim, act_dim, max_length, returns_to_go) and be mindful of offline RL challenges such as data coverage for desired returns.