The Decision Transformer has brought innovation to the field of offline reinforcement learning by reapproaching the trajectory optimization problem from a sequence modeling perspective. This approach enhances system scalability, enables effective cred...
The Decision Transformer has brought innovation to the field of offline reinforcement learning by reapproaching the trajectory optimization problem from a sequence modeling perspective. This approach enhances system scalability, enables effective credit assignment over long-term dependencies, and demonstrates excellent generalization performance across various domains. However, it faces two significant limitations. First, sample-based learning, which relies on offline data, covers only a limited portion of the state-action space, leading to compounding errors when encountering out-of-distribution situations. Second, its autoregressive architecture exacerbates these errors, leading to inaccuracies that accumulate and reduce robustness in long-horizon tasks.
Meanwhile, Flow Matching has emerged as a powerful generative modeling technique. It learns to transfer probability mass from a simple prior distribution to a complex data distribution through a continuous vector field. This process bypasses the need for stochastic sampling and likelihood estimation, enabling faster inference and more stable training while maintaining expressive representational power.
Building on these insights, this paper proposes a novel integrated framework, DT+FM, which combines the strengths of Decision Transformer and Flow Matching. By replacing sample-based learning with distribution-based learning, DT+FM broadens coverage of the state-action space and reduces compounding errors. Experimental results in continuous control environments using MuJoCo and Robomimic show that the proposed DT+FM consistently outperforms the original Decision Transformer, particularly demonstrating greater stability and consistent performance across tasks characterized by long episodes and complex dynamics. These results highlight DT+FM’s promise as a compelling direction for advancing offline reinforcement learning.