VOiLA: Vectorized Online Planning with Learned Diffusion Model for POMDP Agents

The VOiLA framework introduces a method for online planning in POMDPs using learned diffusion models to create task-agnostic POMDP models. It employs conditional diffusion models for transition and observation sampling, which are distilled into compact feedforward generators to facilitate GPU-accelerated planning with the Vectorized Online POMDP Planner (VOPP). Experimental results show that VOiLA reduces sampling costs significantly and outperforms the Recurrent Soft Actor Critic in benchmark tests while demonstrating effective real-world task execution with simulated data.