Researchers led by Tina Lakhani have built a test that shows at which steps of a diffusion model's sampling the random noise shapes the result, according to a paper on arXiv. Diffusion models are used more and more in robotics and control because they can express several workable solutions to a task.

Across five combinations of models and samplers on three robotics benchmarks, the effect of noise differed widely: changes could grow, partly shrink or almost vanish in later steps. The team named the short spans where noise matters most stochastic diversity windows.

They used the windows in two ways: adding more noise inside them to explore more solutions, and a hybrid sampler that skips steps outside them. The hybrid cut the number of network evaluations by up to 40 to 60% and the time taken by up to 85%, with no consistent change in feasibility, cost or diversity on held-out tasks.

Source: arXiv Robotics (cs.RO)

Background

Tina Lakhani

Robotics researcher

Robotics researcher.