Planning under Noise with Neural Operators

Michael Chen; Hong Chul Nam; Xi Deng; Anima Anandkumar

cs.LGOct 20, 2026

Planning under Noise with Neural Operators

Authors: Michael Chen, Hong Chul Nam, Xi Deng, Anima Anandkumar

Organizations: Computing and Mathematical Sciences, California Institute of Technology · Center for Data Science, New York University

Abstract

Planning 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. Building on an established screened-Poisson formulation of stochastic optimal control, we address two challenges: specifying boundary conditions that support navigation across varying geometries and efficiently computing fields with reliable gradients. Zero-Dirichlet obstacle boundaries terminate sampled paths at walls, weakening the goal signal along long routes and making it harder to estimate reliably. We instead impose reflecting Neumann conditions on obstacles and a Dirichlet condition at the goal, allowing sampled paths to continue exploring free space until they reach the target. To solve the resulting PDE without a configuration-space mesh, we extend a Monte Carlo (MC) solver, Walk on Stars (WoSt), to higher dimensions. However, local MC fluctuations can reverse estimated gradient directions and cause planning failures, making direct planning require costly sampling. PUN amortizes this computation by learning geometry-to-field mappings from noisy, partially converged Monte Carlo labels, avoiding expensive convergence for every training example. At inference, test-time scaling uses bounded local search guided by the predicted global field. We evaluate PUN on planning tasks across dimensions. Across four noise levels, PUN matches or outperforms all compared planners, achieving 95-100% success on the 4D replanning benchmark while maintaining the lowest runtime. At the highest tested noise level in 4D, PUN achieves 10.7-60.6% relative gains in success rate over other methods while running 3.29-32.15 times faster.

Figures & tables

Fig. 1: (a) Fast replanning. PUN has the lowest mean runtime on unseen 4D maps: 10.8 ms/start, compared with 32.8–387.1 ms/start for the baselines. Runtimes are averaged equally over the four execution-noise levels in Table 5 and plotted on a logarithmic axis. (b) Resolution transfer. FNO maintains over 90% local-search success across all five resolutions without retraining, reaching 92.50% at $4096\times 4096$ after training at $256\times 256$ . Missing points indicate out-of-memory failures. (c, d) Continued goal seeking. In the illustrated geometry, the solution induced by the Dirichlet boundary cannot propagate goal information further since most trajectories are absorbed by obstacles. Consequently, it is more difficult for planners to find the target than with the mixed boundary condition.

Panel of Figure 2 — Fig. 2: Planning Under Noise: (A) Planning: a neural operator maps geometry, goal, cost, and noise inputs to a desirability field. Grid-based local search selects feasible routes with high endpoint desirability, executes a short prefix, and replans. (B) Training: Walk-on-Stars (WoSt) estimates provide pointwise supervision; neural caching supplies learned continuation at the walk horizon.

Figure 3

Fig. 3: Replanning trajectories on the same 4D manipulator, start, and goal at $\lambda=0.25$ (top) and $\lambda=2$ (bottom). Columns show successive execution states; PUN’s trajectory is orange. PUN reaches the goal in both illustrated rollouts. Aggregate success and runtime are reported in Table 5 .

Model	Gradient (%)	Local search (4) (%)	Local search (10) (%)	Relative $L_{2}$
UPT	2.48	2.78	3.57	0.5533
Transolver	3.26	5.34	15.30	0.5158
WoSt	2.71	6.30	36.45	0.1852
GINO	46.56	51.24	53.11	0.2628
U-Net	74.21	84.22	88.98	0.1605
DISCO	82.49	87.60	91.08	0.1041

Table 1: 2D application-benchmark evaluation of neural backbones trained with WoSt supervision and direct WoSt inference: planning success and relative

L_{2}

error against the FEM ground-truth field.

Table 6

Dim.	Method	$\lambda=0.25$	$\lambda=1$	$\lambda=2$	$\lambda=4$
2D	A*	59.00 / 43.63	42.33 / 47.09	35.00 / 51.18	23.67 / 53.76
	Dijkstra	59.00 / 154.94	42.33 / 174.95	35.00 / 203.41	23.67 / 239.97
	RRT	82.00 / 98.45	68.67 / 103.87	52.33 / 112.27	36.67 / 124.71
	MPPI	84.00 / 1734.11	61.00 / 1773.31	45.67 / 1838.73	31.67 / 2126.12
	PI 2 -CMA	78.33 / 1330.76	60.00 / 1484.91	48.00 / 1524.70	35.00 / 1606.82
	PUN (zero-shot)	91.67 / 9.55	74.00 / 11.14	62.33 / 14.07	45.67 / 14.51

Table 5: Replanning under execution noise on 2D and 3D application benchmarks and 4D manipulators. Each entry reports success (%) / runtime (ms/start). Bold marks the highest success and lowest runtime within each dimension and noise level, including ties.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Definition	Meaning
Control problem
$\xi,\ x$	§ 2.1	starting point of the process or walk; conditioned as $\mid\xi=x$ where the start is not otherwise fixed
$X^{\xi,\mathbf{u}}_{t}$	Eq. 1	controlled process at time $t$ , from the prescribed start $\xi$ under control $\mathbf{u}$
$\mathbf{b}$	Eq. 1	drift of the controlled system ( $\mathbf{b}=0$ in the specialization)
$G,\ R$	Eq. 1 , Eq. 2	control matrix and control penalty ( $G=R=I$ in the specialization)
$\Sigma$	Eq. 1	diffusion tensor; $\Sigma=\lambda I$ when $G=R=I$ under matching

Appendix

Table 6: Notation, with the equation, assumption, or section where each symbol is defined.

$\rho/r$	$0$	$1/6$	$1/3$	$1/2$	$2/3$	$5/6$	$1$
$\kappa r=1$	1.0000	0.9880	0.9593	0.9232	0.8880	0.8614	0.8509
$\kappa r=4$	1.0000	0.8558	0.6157	0.4086	0.2629	0.1766	0.1466

Appendix

Table 7: Star-step weight

Q^{\kappa}_{3}(\rho,r)

r=1

. The weight is

1

for a zero-length step and decreases with the realized hit distance

\rho

, reaching

\gamma_{3}(\kappa r)=\kappa r/\sinh(\kappa r)

\rho=r

. A step cut short by an obstacle loses less weight than a full sphere step.

$d$	$G_{\mathbb{R}^{d}}(r)$	$G^{D}_{B_{R}}(r)$	$\gamma_{d}(z)$
$2$	$\dfrac{K_{0}(\kappa r)}{2\pi}$	$\dfrac{1}{2\pi}\!\left[K_{0}(\kappa r)-\dfrac{K_{0}(\kappa R)}{I_{0}(\kappa R)}I_{0}(\kappa r)\right]$	$\dfrac{1}{I_{0}(z)}$
$3$	$\dfrac{e^{-\kappa r}}{4\pi r}$	$\dfrac{1}{4\pi r}\!\left[e^{-\kappa r}-e^{-\kappa R}\dfrac{\sinh(\kappa r)}{\sinh(\kappa R)}\right]$	$\dfrac{z}{\sinh z}$
general	$\dfrac{\kappa^{\nu}}{(2\pi)^{d/2}}\dfrac{K_{\nu}(\kappa r)}{r^{\nu}}$	$\dfrac{\kappa^{\nu}}{(2\pi)^{d/2}r^{\nu}}\!\left[K_{\nu}(\kappa r)-\dfrac{K_{\nu}(\kappa R)}{I_{\nu}(\kappa R)}I_{\nu}(\kappa r)\right]$	$\dfrac{(z/2)^{\nu}}{\Gamma(\nu+1)I_{\nu}(z)}$

Appendix

Table 8: Screened Poisson kernels by dimension. All entries follow from Eq. 17 , Eq. 18 , and Eqs. 20 – 21 with

\nu=d/2-1

Dim.	Method	$\lambda=0.25$	$\lambda=1$	$\lambda=2$	$\lambda=4$
2D	A*	[55.92, 62.08]	[39.16, 45.50]	[31.90, 38.10]	[20.89, 26.45]
	Dijkstra	[55.81, 62.19]	[39.15, 45.51]	[31.99, 38.01]	[21.08, 26.26]
	RRT	[79.37, 84.63]	[65.68, 71.66]	[49.14, 55.52]	[33.46, 39.88]
	MPPI	[81.70, 86.30]	[57.77, 64.23]	[42.45, 48.89]	[28.82, 34.52]
	PI 2 -CMA	[75.76, 80.90]	[56.81, 63.19]	[44.76, 51.24]	[31.95, 38.05]
	PUN (zero-shot)	[89.90, 93.44]	[71.27, 76.73]	[59.15, 65.51]	[42.55, 48.79]

Appendix

Table 9: 95% confidence intervals for replanning success rates (%) corresponding to Table 5 .

Dim.	Method	$\lambda=0.25$	$\lambda=1$	$\lambda=2$	$\lambda=4$
2D	A*	[24.99, 62.27]	[26.25, 67.93]	[28.10, 74.26]	[29.44, 78.08]
	Dijkstra	[79.37, 230.51]	[88.65, 261.25]	[102.34, 304.48]	[123.40, 356.54]
	RRT	[69.41, 127.49]	[72.10, 135.64]	[78.18, 146.36]	[82.46, 166.96]
	MPPI	[1227.92, 2240.30]	[1257.88, 2288.74]	[1279.80, 2397.66]	[1400.54, 2851.70]
	PI 2 -CMA	[964.06, 1697.46]	[1053.67, 1916.15]	[1067.48, 1981.92]	[1079.98, 2133.66]
	PUN (zero-shot)	[7.79, 11.31]	[8.80, 13.48]	[10.71, 17.43]	[10.19, 18.83]

Appendix

Table 10: 95% confidence intervals for replanning runtime (ms/start) corresponding to Table 5 .

Test set	Gradient	Local search (10)
3D application benchmarks	65.95	89.46
4D manipulators	97.50	99.48

Appendix

Table 11: FNO (WoSt) planning success (%) with gradient following and local search.

Method	2D zero-shot	4D manipulators
WoSt + FNO	1.20	2.79
Direct WoSt	1128.60	160.59

Appendix

Table 12: Full test-time cost (ms/start) on the 2D zero-shot application-inference subset and 4D manipulator test set. Both methods use local search in 2D and in 4D.

Fig. 4: Zero-shot FNO field predictions and extracted paths at $128\times 128$ , $200\times 200$ , and $256\times 256$ (left to right). Color shows predicted desirability $\Psi_{\theta}$ , gray denotes obstacles, circles mark starts, and the star marks the shared goal in each map.

Panel of Figure 16 — Fig. 5: Goal-conditioned field prediction and reuse across initial states.

Panel of Figure 17 — Fig. 6: WoSt walk and label statistics: (a) walk-length variability and boundary treatment; (b) label and gradient variability on a fixed map, with differences measured against a finite Monte Carlo reference; (c) spatial variation in mean walk length.

Fig. 7: Local search on one fixed urban-map field: (A) full map and routes; (B,C) enlarged neighborhoods for local search (4) and (10). Greater depth exposes endpoints farther along the passage. Gray cells are obstacles; overlays mark candidate regions and routes. Table 1 reports aggregate controller success.

Motion Planning

Motion planning finds safe, feasible paths and trajectories for robots and vehicles amid geometric, dynamic, and task constraints.

See all 405 papers

Jun 2, 2026

Neural Navigation Functions for Zero-Shot Generalizable Motion Planning

We introduce Neural Navigation Functions (Neural-NF), a learned reactive navigation function capable of zero-shot transfer across unseen environment geometries. Neural-NF places data-driven adaptation within a structured elliptic planner, where the navigation objective is learned while planner structure is preserved by construction. Specifically, intrinsic Laplacian-derived features are mapped to local PDE coefficients, and solving the resulting boundary value problem produces a globally consistent value function on each target domain. For every admissible learned model, the resulting policy is collision-free, provides monotonic descent and a global minimum at the goal by construction. This admits a linearly-solvable optimal-control interpretation for any parameter setting. Empirically, Neural-NF achieves strong zero-shot transfer across diverse geometries and outperforms learned planners that directly predict the value function by up to a

5\times

improvement.

Benjamin D. Shaffer, Pei-An Hsieh, Brooks Kinch +2

University of Pennsylvania, United States 1Department of Mechanical Engineering and Applied Mechanics · University of Pennsylvania, United States 2Department of Electrical and Systems Engineering

Zero-Shot Learning Roadmap-Based Motion Planning Safe Motion Planning Motion Planning Robot Motion Planning

Ask PDF

Oct 8, 2026

Real-Time Motion Planning with Dynamic Hazards: Classical vs. Learning-Based Methods

We study real-time motion planning in dynamic hazard fields through a controlled comparison between classical planning and learning-based methods. Rather than introducing a new planner, we construct a unified benchmark in which representative classical and learning-based methods face the same environments, motion constraints, information assumptions, and evaluation metrics. The test environment consists of planar domains populated with rotating sprinkler-like hazards that generate time-varying forbidden regions via sweeping angular sectors. Our results show a clear regime shift. In deterministic environments, classical planners achieve near-perfect success and higher-quality paths, though sometimes at the cost of substantial planning or replanning time. Under stochastic obstacle dynamics, however, online search becomes strongly budget-sensitive: low budgets lead to frequent failure, while high budgets improve success at the cost of latency and longer trajectories. PPO-based policies, trained under the same scenario distribution, consistently outperform in latency, success rate, and path quality in these stochastic regimes. Overall, the results indicate that uncertainty in obstacle evolution, more than partial observability, is the dominant factor determining which planning paradigm is practically effective for the problem at hand.

Eran Iceland, Alexander Tuisov, Oren Gal +2

School of Engineering and Computer Science, The Hebrew University of Jerusalem, Jerusalem, Israel · Faculty of Data and Decision Science, Technion Israeli Institute of Technology, Haifa, Israel · Hatter Department of Marine Technologies, University of Haifa, Haifa, Israel +1

Motion Planning Dynamic Obstacle Avoidance Robot Policy Learning

Ask PDF

Sep 29, 2026

Learning to Plan from Random Exploration

Random exploration reveals how an environment can be traversed before a goal is specified. Can this experience support long-range planning without policy-improvement training? Our random-walk analysis explains what temporal relations contain: short horizons reveal geodesic geometry in the diffusion limit, while longer horizons reveal connectivity between regions before mixing removes these distinctions. We learn these relations with a conditional energy-based model that estimates temporal log-density ratios through horizon-conditioned embeddings. The model is trained on observation pairs by noise-contrastive estimation, without action or reward labels. The planner queries these learned relations at different horizons as it moves toward the goal. At test time, a separate local dynamics model predicts candidate action outcomes, and the temporal model evaluates their progress toward the goal by selecting or aggregating estimated improvements across horizons. The agent executes one action and replans with both models fixed. Experiments demonstrate long-range maze planning from random exploration using states and images. Learned score fields, embedding probes, and planned routes exhibit properties of a multiscale cognitive map. We further demonstrate egocentric navigation from random exploration and manipulation planning from suboptimal data.

Deqian Kong, Guangyan Sun, Sheng Cheng +6

UCLA · University of Minnesota · Amazon AGI +3

Long-Horizon Planning Model-Based Planning Long-Horizon Reasoning

Ask PDF

Abstract

Motion Planning

Explore similar work

Neural Navigation Functions for Zero-Shot Generalizable Motion Planning

Real-Time Motion Planning with Dynamic Hazards: Classical vs. Learning-Based Methods

Learning to Plan from Random Exploration