Navigating Uncertainties with Active Inference and Probabilistic Diffusion

This paper presents a novel approach of integrating Active Inference (AIF) and Probabilistic Diffusion (PD) for autonomous navigation. AIF is a promising paradigm for “optimal” control without the need for complex and extensive dynamical models, especially for non-linear and time-variant systems. On the other hand, PD can be used to generate navigational strategies (e.g., prediction of motions) for reaching a target. When combined, the two form a locally optimal navigator, where AIF optimizes the future strategies sampled from PD, while the samples are adjusted using the variational free energy from AIF when new actual observations are made. We demonstrate the framework in a simulated parking lot scenario and highlight two key advantages over reinforcement learning (RL): learning without handcrafted rewards and improved generalization to novel settings. Though AIF and RL differ fundamentally, our benchmark shows that AIF+PD achieves comparable performance in complex multi-agent parking tasks.