DASC 2026 @ Orlando, Florida

Runtime Safety Filtering for Learned Small UAS Separation Policies under GNSS Degradation

Action filtering vs. observation filtering

Alex Zongo, Peng Wei

Department of Mechanical and Aerospace Engineering · The George Washington University

Preprint Slides Cite
  • 3rd Best of Conference
  • Best of Air Traffic Management Track
  • Best of UAM/UAS Traffic Management and Separation Assurance Session

Where should a safety filter intervene?

Correct what the policy sees, or constrain what it does? Two runtime architectures share a worst-case state estimate, but produce very different safety outcomes.

Action filtering vs. observation filtering (schematic)A degraded observation feeds a worst-case state estimate. In action filtering, the policy acts on the degraded observation and a barrier check corrects its action. In observation filtering, the estimate is fed to the policy as a corrected input. observation õ (degraded) worst-case estimate ŝ* Action filtering · constrain what it does π(õ) a barrier checkh(ŝ*, a) ≥ 0 ? asafe Observation filtering · correct what it sees π(ŝ*) a
Schematic. One worst-case estimate, two interventions: action filtering corrects the output, observation filtering corrects the input. Neither retrains the deployed policy.
fewer near mid-air collisions
90%
NMACs per episode
0.24 → 0.02
average separation
250 → 375 m

Reported simulation comparison in the DASC presentation (slide 10); Rmax = 0.7 and α = 5 s. These are experimental results, not a general safety guarantee.

01 / The problem

A safe decision needs a trustworthy observation.

A decentralized separation policy uses nearby aircraft positions, velocities, and headings to select a speed change. GNSS multipath, blockage, or interference can distort those observations: an apparently clear path may conceal a closing conflict. The policy acts on the reported state, while safety is measured on the true state.

The controller

Three speed actions: slow down, hold speed, or speed up. Heading follows the structured route.

The safety measure

A near mid-air collision (NMAC) occurs when two aircraft pass within the 100 m separation threshold.

The uncertainty

Bounded adversarial observation degradation, with contamination rate R and a design bound Rmax selected before deployment.

02 / The comparison

Same estimate. Different decision authority.

Both architectures estimate the worst-case traffic state under the assumed uncertainty bound. What changes is how that estimate interacts with the learned policy.

Action filtering

Enumerate candidate speed actions, test a discrete-time control barrier function condition, and keep the feasible action closest to the policy's proposal.

Observation filtering

Construct a corrected observation from the worst-case estimate, pass it to the policy, and retain the policy's chosen action.

Action filtering enumerates and tests actions against a barrier condition. Observation filtering corrects the input and preserves the policy output.
The barrier condition can veto an action only in the action-filtering architecture. Slide 8

03 / Experimental evidence

Correcting the input improves separation.

In the reported comparison, observation filtering reduces NMACs and increases average separation. Action filtering remains close to the unfiltered baseline. The curves below show how both measures vary as the actual contamination rate increases.

NMACs and average separation versus contamination rate. Observation filtering produces fewer NMACs and greater separation than action filtering and the nominal policy.
Nominal policy, action filter, and observation filter at Rmax = 0.7 and α = 5 s. Shaded bands show standard error. Slide 10
How much does the assumed degradation bound matter?

Smaller assumed bounds provide limited benefit. Larger bounds surface more conservative traffic states and improve separation in these experiments. The filter is not given the true contamination rate.

Observation-filter performance for assumed contamination bounds from 0.1 to 1.0. Larger bounds reduce NMACs in the evaluated scenarios.
Sensitivity to the design bound Rmax. Slide 11
How sensitive are the filters to the barrier parameter?

Increasing α changes the balance between separation distance and closing rate in the barrier function. Action filtering is sensitive to that choice; the observation-filter curves remain comparatively similar.

Side-by-side parameter sweeps: action filtering degrades as α grows, while observation filtering maintains low NMAC rates. The panels use different vertical scales.
Barrier-parameter sensitivity. Note the different vertical scales for action and observation filtering. Slide 12

04 / Interpretation & scope

Preserve the policy's learned safety behavior.

The presentation offers a possible explanation: the barrier tests a one-step condition, while the trained policy resolves encounters over multiple steps. Supplying a conservative observation lets the policy use that learned behavior. This is an interpretation of the observed results, rather than a universal ranking of safety-filter architectures.

A policy-dependent finding

The benefit depends on how the deployed policy was trained and the safety behavior it learned.

A cost in conservatism

A more cautious observation may trigger earlier or more frequent slowing down. Safety and efficiency must be assessed together.

Simulation scope

These results concern the evaluated learned policy, structured routes, and bounded GNSS-degradation model. They do not establish flight certification or hardware validation.

Reference

Cite this work

Zongo, A., & Wei, P. (2026). "Runtime Safety Filtering for Learned Small UAS Separation Policies under GNSS Degradation." arXiv:2607.10014. Accepted at DASC 2026.

BibTeX 
@misc{zongo2026runtime,
  title = {Runtime Safety Filtering for Learned Small UAS Separation
           Policies under GNSS Degradation},
  author = {Alex Zongo and Peng Wei},
  year = {2026},
  eprint = {2607.10014},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  note = {Accepted at DASC 2026},
  url = {https://arxiv.org/abs/2607.10014}
}