PUN
A planning framework based on neural operators that learns a noise-aware value field from noisy Monte Carlo labels to enable efficient planning.
Timeline (last 90 days)
PUN (Planning Under Noise) is a neural operator-based planning framework designed to learn a noise-aware value field from noisy Monte Carlo labels, thereby achieving efficient planning in unseen geometric environments.
PUN is built on the screened-Poisson equation formulation of stochastic optimal control.
The PUN authors instead impose a reflecting Neumann condition on obstacles and a Dirichlet condition at the goal point, allowing sampled paths to continue exploring free space until reaching the goal, significantly improving signal propagation in long-distance navigation.
By learning the mapping from geometry to value field from noisy, partially converged Monte Carlo labels, PUN amortizes computation into the training phase, avoiding expensive convergence for every sample.
At the core of PUN is a neural operator that takes in geometric information and outputs a noise-aware value field.
PUN's training data is generated by noisy Monte Carlo simulations that have not fully converged, so the labels themselves contain fluctuations; the model learns to extract robust field representations from these imperfect labels.
At inference time, PUN employs test-time scaling: it uses the predicted global field to guide a bounded local search, thereby improving planning quality while maintaining computational efficiency.
The research evaluates PUN on planning tasks across multiple dimensions, covering four noise levels.
On the 4D replanning benchmark, PUN maintained the lowest runtime at a success rate of 95-100%.
At the highest noise level in 4D, compared with other methods, PUN achieved a relative improvement in success rate of 10.7-60.6% while running 3.29-32.15 times faster.
2 more timeline events
All Facts (12)
PUN (Planning Under Noise) is a neural operator-based planning framework designed to learn a noise-aware value field from noisy Monte Carlo labels, thereby achieving efficient planning in unseen geometric environments.
50%UnverifiedPUN is built on the screened-Poisson equation formulation of stochastic optimal control.
50%UnverifiedThe PUN authors instead impose a reflecting Neumann condition on obstacles and a Dirichlet condition at the goal point, allowing sampled paths to continue exploring free space until reaching the goal, significantly improving signal propagation in long-distance navigation.
50%UnverifiedBy learning the mapping from geometry to value field from noisy, partially converged Monte Carlo labels, PUN amortizes computation into the training phase, avoiding expensive convergence for every sample.
50%UnverifiedAt the core of PUN is a neural operator that takes in geometric information and outputs a noise-aware value field.
50%UnverifiedPUN's training data is generated by noisy Monte Carlo simulations that have not fully converged, so the labels themselves contain fluctuations; the model learns to extract robust field representations from these imperfect labels.
50%UnverifiedAt inference time, PUN employs test-time scaling: it uses the predicted global field to guide a bounded local search, thereby improving planning quality while maintaining computational efficiency.
50%UnverifiedThe research evaluates PUN on planning tasks across multiple dimensions, covering four noise levels.
50%UnverifiedOn the 4D replanning benchmark, PUN maintained the lowest runtime at a success rate of 95-100%.
50%UnverifiedAt the highest noise level in 4D, compared with other methods, PUN achieved a relative improvement in success rate of 10.7-60.6% while running 3.29-32.15 times faster.
50%UnverifiedPUN's contribution lies in combining neural operators with the PDE formulation of stochastic optimal control, achieving fast noise-aware planning in unknown geometries by improving boundary conditions and amortizing Monte Carlo computation.
50%UnverifiedPUN's test-time scaling strategy offers learned planners an approach that balances global guidance with local refinement.
50%