Skip to main content

Sage-HaLow — First-Pass Summary

Authors: Ben Owusu-Amo and Hairik Honarchian Saki (Colorado State University)

Project: A battery-powered ESP32-S3 camera node that reaches the Sage data repository over Wi-Fi HaLow (802.11ah), at sites with no wired power and no usable Wi-Fi. The node scores every frame on-device and only spends radio time when the scene actually changed. Short-term goal: images and telemetry landing in Beehive from a real deployment, plus a measured answer to "how much of what we send is junk" — before any model gets built.

Sage Summer Camp 2026 · University of Illinois Chicago, Electronic Visualization Laboratory (EVL).

Hardware​

  • Node: Heltec HT-HC33 — ESP32-S3 plus a Morse Micro sub-GHz radio on one board. Camera, microSD, battery with a solar input.
  • Gateway: Heltec HT-H7608 Wi-Fi HaLow gateway.
  • Edge: Waggle/Sage node H03D (sgt-thor-1423325056007-H03D), which runs the bridge plugin.
  • Camera enclosure provided by Pete for Summer Camp evaluation.

Firmware​

Each wake: capture → score → decide → (maybe) upload → deep sleep.

  • Scene score: grid-downscaled grayscale frame diff against a stored reference, with a lighting-shift rejection path so sun and cloud changes do not trigger it.
  • State machine: BASELINE → AROUSED → ACTIVE → COOLDOWN, with hysteresis, an N-frame confirm so one blip cannot trip it, and exponential backoff. The reference frame and FSM state persist in RTC memory and on SD across deep sleep.
  • Energy governor: SoC curve, charge-rate trend, and battery bands with hysteresis cap what the FSM is allowed to do. Composition is FSM → operator override → energy cap → charge-rate promotion. 58 host tests.
  • Transport: MQTT. ACK before delete (the SD copy survives until the base station ACKs), 8 KB chunking for QXGA frames, and config pushed down under a timestamp + expiry lease so a stale retained message cannot replay.
  • Validated on hardware 2026-07-23: a live scene change drove BASELINE → AROUSED → ACTIVE → publish, end to end.

Sage integration​

plugin-halow-camera is the base-station receiver repackaged as a Waggle edge app, running on node H03D. It subscribes to the broker, reassembles chunked JPEGs, embeds EXIF, and calls plugin.upload_file() instead of writing to the node's disk. The scene heartbeat becomes measurements: halow.camera.scene.score, .scene.state, .sleep.seconds, .resolution, .image.bytes, .image.complete.

Timestamps are nanoseconds. When a node reports time_synced=true the node's own capture time is used, not the receive time — so Beehive indexes an image at the moment it was taken, and a "last hour" query will miss a frame captured five hours ago.

Result: images and telemetry are retrievable from the Sage data repository.

https://storage.sagecontinuum.org/api/v1/data/Pluginctl/sage-halow-bridge-latest/00004cbb4701cbaf/1785251056000000000-20260728_150416_B43A45A45644.jpg

A 2048×1536 QXGA capture from camera node B43A45A45644, uploaded by the plugin from node H03D and stored by Beehive; 80,979 bytes. Opening the object requires Sage credentials (an unauthenticated GET returns 401), though the measurement record itself is public through the query API. Pod halow-bridge in Running. Built on waggle/plugin-base:1.1.1-base (~950 MB, no CUDA) — the -ml base unpacks to roughly 25 GB and is the known cause of Pending pods on this node.

Caveat: the frames that have crossed the plugin so far are genuine camera captures (verified pixel-identical to 20260723_220643_B43A45A45644.jpg), but they were replayed onto the broker by a test publisher under the IDs SYNTHA–SYNTHD rather than published live by the ESP32 over HaLow in the same run. The four cases were chosen to exercise every path: single-message upload, 5-chunk QXGA reassembly, a truncated _PARTIAL, and a _CLOCKUNSET node. What is still untested is one continuous live run, ESP32 → gateway → broker → plugin → Beehive.

Not built: the model​

The last stage does not exist yet. The target class decides the model (animal, smoke, and snow line are different problems with different training data and different costs of being wrong), decides what counts as junk in the metric above, and decides where inference runs — a small suppressor on the ESP32 versus a real classifier on the Waggle node's GPU, which sets the flash, RAM, and latency budgets. Building it before the deployment is defined means optimizing against a number nobody has measured.

Next​

  1. Deploy at a site with real traffic; let the plugin carry the frames.
  2. Pull decisions.csv off the SD card, label the batch with MegaDetector or some other AI identification model.
  3. Re-read the gate for an actual GO / NO-GO.
  4. Validate the energy governor on real battery and solar hardware — it has only ever run against a simulated state of charge.

Acknowledgments​

Firmware for the HT-HC33 platform by Ben Owusu-Amo. Heltec HT-HC33 hardware and camera enclosure provided by Pete for evaluation during the Summer Camp. This work was supported in part by the National Science Foundation under Awards No. 2331263 and 2436842.