Bayesian Flow Networks Enhance Offline Trajectory Planning for Robotics.

Offline reinforcement learning (RL) is crucial for developing decision policies from existing datasets without needing real-time interaction with an environment. Current methods often rely on continuous diffusion models for generating trajectories, which can be challenging to adapt for discrete planning tasks that require categorical formulations. A new framework, BFN-RL, addresses this by leveraging Bayesian Flow Networks (BFNs) to create a unified generative modeling approach for offline RL. Unlike diffusion models that evolve noisy data instances, BFN-RL iteratively adjusts distribution parameters, allowing it to naturally handle both discrete and continuous trajectory spaces within a single probabilistic model. The framework includes a categorical planner for generating future state sequences and an inverse-dynamics model to translate these states into actions. Evaluations across discrete planning and continuous control scenarios demonstrate BFN-RL's ability to generate effective trajectories, establishing BFNs as a versatile foundation for offline trajectory planning across different data types.