Disturbance-Aware Informative Planning
in Spatiotemporal Environments

Kaleb Ben Naveed1Haejoon Lee1Anouck Girard2Dimitra Panagou1,3

1 Department of Robotics, University of Michigan, Ann Arbor2 Embry-Riddle Aeronautical University3 Department of Aerospace Engineering, University of Michigan, Ann Arbor

arXiv Coming soon Code Slides

Learn wind where it matters
for monitoring.

Robots performing persistent monitoring must revisit regions as information decays. Wind can delay or deflect a planned route, so a short path can deliver poor monitoring performance.

We use geometric alternatives and predicted clarity deficits to identify where wind knowledge matters. The robot first gathers measurements along relevant paths. Once weighted uncertainty is sufficiently low, disturbance-aware model-based diffusion (MBD) plans monitoring trajectories by predicting motion and sensing under sampled wind fields.

From useful paths to informed motion

Sensing continues during both wind learning and monitoring.

Five-step method: geometric paths, relevant wind uncertainty, targeted measurements, disturbance-aware optimization, and execution.
Clarity deficitWind uncertaintyRelevant pathsSelected trajectory
Task and geometric paths

Predict future clarity deficits and retain routes through different passages.

Conceptual illustration. Blue predictions show how wind changes motion; the teal alternative is selected for lower expected monitoring cost. View the animated GIF ↗

Experiments

Recorded simulation results

Five methods in each environment. The shared comparison time includes wind learning; use the GIF or slider controls to inspect each run.

Loading simulations…

The same recordings, in 3D

Drone inspection environments

The saved planar trajectories are replayed at a fixed display height. Moving tracers use each scenario’s stationary CFD wind field; the architecture is illustrative.

Scenario 1 · Drone inspection courtyard

The recorded trajectories in an industrial courtyard, with moving tracers from the saved stationary wind field. Select any method and replay its wind-learning and monitoring stages.

Open the 3D view in a full window ↗

Scenario 2 · Industrial courtyard · opposing wind

The same courtyard layout as Scenario 1, with the recorded strong wind in both the central and lower passages. The upper route remains sheltered.

Open the 3D view in a full window ↗

Scenario 3 · Sheltered inspection plaza

A glazed pavilion sits within the upper monitoring plaza. The recorded strong wind lies in the lower-right forecourt, away from the main monitoring region.

Open the 3D view in a full window ↗

Scenario 4 · Utility campus · selective wind learning

Workshops and storage tanks form a utility campus. The monitoring bays sit on the west side, while the larger eastern service yard adds space that need not be fully surveyed.

Open the 3D view in a full window ↗