Do Confidence Gates Survive Structured Communication Loss? A Frozen-Policy Audit of Residual Multi-UAV Coordination
Abstract Confidence signals are widely used to decide when a learned residual policy should refine a nominal controller, but in networked teams the signal is itself produced by the communication process. We audit this execution-time design choice in simulated multi-UAV maritime search and tracking, where a sigmoid gate scales a 450-parameter residual by post-fusion belief confidence. Using six independently trained gated/always-on policy pairs with frozen weights, evaluated on 23 held-out seeds, the gate's advantage over an always-on residual under a nominal disc channel is small (pooled margin +0.009, 95% CI [-0.002, +0.020]). Zero-shot transfer to a path-loss and co-channel-interference channel lowers the margin in all six pairs (mean shift -0.026, CI [-0.048, -0.005]) and makes it negative on average (-0.017), whereas the always-on residual shifts by only -0.006. Counterfactual interventions show that the mechanism is pair-dependent: replacing only the scalar trigger or opening the gate removes the deficit in two pairs but not in the others. In the case-study pair, fidelity-matched i.i.d. trigger corruption predicts a margin of -0.003 against the observed -0.036, whereas persistent per-UAV dropout and a shared cross-UAV outage regime each reproduce about 70% of the excess. Training on the structured channel recovers most of the loss in that pair but does not surpass the always-on residual. Gate evaluations should therefore include an always-on control and temporally persistent, spatially correlated link loss.
Authors
- Manisha Khulbe (ORCID: https://orcid.org/0000-0002-6114-8303)
- Satish Mishra
Institutions
- Netaji Subhas University of Technology (IN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-17
- DOI
- https://doi.org/10.21203/rs.3.rs-11729253/v1
- Primary Topic
- Reinforcement Learning in Robotics
- Type
- preprint