PUN neural operator plans under control noise in unseen geometries
Original title: Planning under Noise with Neural Operators
On a 4D replanning benchmark it reports 95-100% success across four noise levels and the lowest runtime among compared planners.
Planning Under Noise (PUN) learns noise-aware value fields for planning in unseen geometries from a screened-Poisson formulation of stochastic optimal control. It uses reflecting Neumann obstacle boundaries, a Dirichlet goal condition, and extends Walk on Stars Monte Carlo solving to higher dimensions. Authors say PUN matches or beats compared planners and, at the highest tested 4D noise, gains 10.7-60.6% relative success while running 3.29-32.15 times faster.
Unconfirmed: Performance figures are the authors' benchmark results; independent verification is not stated.
- Training uses noisy, partially converged Monte Carlo labels rather than full convergence per example.
- Inference uses bounded local search guided by the predicted global field.
- Highest-noise 4D result shows 10.7-60.6% relative success gains.
- That setting is reported 3.29-32.15 times faster than other methods.
From the original article
1 min read at arxiv.orgPlanning under uncertain control execution noise requires reliable feedback policies across diverse environments. We introduce Planning Under Noise (PUN), a neural operator that learns noise-aware value fields for planning in unseen geometries.
By Michael Chen · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.