A research paper out of arXiv proposes Valdi, a system that tries to make diffusion models useful for real-time robotic planning — something the field has struggled with.
Diffusion models are good at representing uncertainty, which makes them appealing for predicting how an environment might evolve. The problem is they are slow: generating a prediction requires many iterative steps, which rules them out for low-latency control loops. Valdi's answer is to collapse that to a single diffusion step at both training and inference, pairing that with a latent-space dynamics model trained end-to-end for Model Predictive Control. In preliminary tests on the CarRacing benchmark, the single-step Valdi matched a plain deterministic neural network baseline.
That result is more cautious than it sounds. Matching a simple baseline is not beating it, and the authors themselves flag a core tension: the richer the model's ability to represent multiple possible futures, the harder it becomes to extract a clean control signal. That tradeoff is the real finding here, not the benchmark score.
Diffusion-based planning is an active research front — several groups are racing to make these models fast enough for real hardware. Valdi is a preliminary step, and the authors say so plainly. The code is public, which at least lets others stress-test those tradeoffs.