Publications Conference paper · Satellite IoT · 2026
To Transmit or Not: Tiny On-Device Transformers for Uplink Prediction in Direct-to-Satellite IoT
IEEE Latin-American Conference on Communications (LATINCOM), 2026
Abstract
Direct-to-Satellite IoT (DtS-IoT) promises to connect massive fleets of devices in regions where terrestrial infrastructure will never reach. Its weakest link, however, is the device itself: DtS-IoT nodes are inexpensive, energy-constrained, and effectively blind, since downlink feedback from sporadically passing Low-Earth Orbit (LEO) satellites is broadcast, slow, and costly, so in practice devices cannot expect their transmissions to be acknowledged. The satellite cannot micromanage its fleet: every transmission decision must be made on-device, based solely on local information, and every wasted uplink drains energy and fuels collisions. We argue that these decisions can be learned, provided the models are small enough to run on highly constrained hardware and to be updated over a scarce downlink. This paper presents a tiny encoder-only Transformer (14k parameters, 64 KB) that predicts, from the device's recent transmission context, whether an uplink will be received. To address the scarcity of DtS-IoT data, we generate synthetic datasets using the open-source FLoRaSat simulator and introduce an ML Sandbox module that embeds trained models directly into the simulation for closed-loop evaluation. Deployed in that closed loop, the model delivers 16.1% more unique packets than an unguided baseline while sending 41.6% fewer transmissions, and its operating frontier dominates a Bidirectional Long Short-Term Memory (Bi-LSTM) of comparable footprint, an analytical link-budget policy, and the blind repetition budget itself: spending the same energy without the model delivers 2.3× fewer packets. Both gains are the currency that matters for blind, battery-powered devices: more data through, less energy spent.
Cite
BibTeX
@inproceedings{choquenaira2026totransmit,
author = {Choquenaira-Florez, Alexander Y. and Fraire, Juan A. and Rivano, Herv{\'e}},
title = {To Transmit or Not: Tiny On-Device Transformers for Uplink Prediction in Direct-to-Satellite {IoT}},
booktitle = {IEEE Latin-American Conference on Communications (LATINCOM)},
year = {2026}
}