Benchmarking Generative Trajectory Models for Active-Inference Control

This research introduces GenAIF, a generative active-inference control framework that uses trajectory models to control complex systems. The study evaluates diffusion models, autoregressive Transformers, CVA, and flow matching in a MuJoCo manipulation task, finding diffusion models to deliver the strongest control while CVAEs offer faster inference. The results emphasize the importance of correct conditioning and trajectory reuse for improved performance and computational efficiency.