Preprint reports faster wireless perception for resource-limited drones

Researchers report that FreshLatent improved a drone-oriented vision and language system's perception when its intermediate data was sent over a noisy wireless link. The adapter trains a small encoder and decoder to handle transmission errors while leaving the larger model unchanged. In tests at 0 dB under the tightest communication budget, it improved two perception measures by 20.79 and 20.87 points over compression trained on clean data. On a Jetson AGX Xavier running in 10-W mode, its encoder used 7.7-9.9x less time and 8.8-10.0x less energy than a heavier codec. The results are from evaluated conditions, not a field deployment.