Planning under uncertainty in partially observable environments typically requires a faithful POMDP model, which is hard to obtain for real-world robots. The VOiLA framework instead learns the transition and observation models from data using conditional diffusion models, and uses learned observation-likelihood models for particle-based belief updates.

To make online planning tractable, the authors distill the diffusion samplers into compact feedforward generators and integrate them with VOPP, a GPU-parallelized online POMDP planner. The distillation reduces sampling cost by up to nearly three orders of magnitude, making learned generative models practical for online use.

Across three benchmark problems, VOiLA performed equal to or better than recurrent RL methods such as Recurrent Soft Actor Critic and DreamerV3, while using less than 10% of the training data and generalizing better to unseen environment configurations. In a physical robot experiment, it completed the task in all ten runs using models trained purely on simulated data.