What can I use only one ESP32-S3 DevKitC to do?

[dabaosun]

I'm using only one ESP32-S3 DevKiC to evaluate the project. From the http://localhost:3000/ui/index.html, seem the 'Live Demo' and 'Sensing' all not expected. For 'Live Human Pose Detection', it always shows 2 persons, in fact, only 1 person there. For 'Sensing', the classifction always switches in 'PRESENT_MOVING', ‘PRESENT_STILL’, ‘ABSENT’.

So, I'm wondering if the ESP32 mesh is neccessary for all fuctions? and Is this project really deployable and usable in a production environment?

[ruvnet]

With a single ESP32-S3 you get:

Works well: Presence detection, motion detection, breathing rate (stationary, ~1-2m), fall detection

Limited: Pose estimation needs 3+ nodes; person counting unreliable with 1 node

The classification switching PRESENT_MOVING/PRESENT_STILL/ABSENT is correct — that is the on-device presence detector working. To get pose estimation, add 2 more nodes (~$18 total) and set up a TDM mesh.

Update to v0.4.3.1 for latest fixes.

[ajonnet]

How we can check for breathing rate?

I have a single esp32 setup, on Web UI it shows connected, but I was unable see breathing rate being displayed anywhere.

Though I was able to fetch it as json via api call but confidence remains very low around 0.32, while I sitting just next to esp node.

[UMER-Devtechs]

So I've tested this exact setup with a single ESP32-S3 DevKitC and ran into the same issues you're describing. The "always shows 2 persons" bug is actually a known limitation — the pretrained model they're using has a fixed output that defaults to 2 unless you're running the full mesh with multiple nodes providing different angles. The classification constantly flipping between PRESENT_MOVING, PRESENT_STILL, and ABSENT tells me the model has no confidence and is just guessing based on noise, which is pretty common when you're only feeding it RSSI data instead of proper CSI. To answer your questions directly: No, the ESP32 mesh isn't strictly necessary for basic presence detection, but from what I've seen in the codebase, most of the "smart" features like accurate people counting and pose detection were built assuming multiple perspectives. And honestly, for a production environment? I wouldn't trust this as-is — between the inconsistent classification, the fake-looking person count, and the lack of real-world validation from other users, this feels more like a research prototype or a demo than something you'd deploy in an actual home or office. If you just need basic presence, you'd be better off with a PIR sensor or mmWave radar that actually works out of the box.

[ruvnet]

A single ESP32-S3 DevKitC isn’t going to work. Which is also mention in the first paragraph of the readme. Use an array 3 or more of esp32 c6 for anything decent. I also use 60ghz mmWave @ruv

…

On Fri, May 15, 2026 at 3:30 PM Umer Nazakat ***@***.***> wrote: So I've tested this exact setup with a single ESP32-S3 DevKitC and ran into the same issues you're describing. The "always shows 2 persons" bug is actually a known limitation — the pretrained model they're using has a fixed output that defaults to 2 unless you're running the full mesh with multiple nodes providing different angles. The classification constantly flipping between PRESENT_MOVING, PRESENT_STILL, and ABSENT tells me the model has no confidence and is just guessing based on noise, which is pretty common when you're only feeding it RSSI data instead of proper CSI. To answer your questions directly: No, the ESP32 mesh isn't strictly necessary for basic presence detection, but from what I've seen in the codebase, most of the "smart" features like accurate people counting and pose detection were built assuming multiple perspectives. And honestly, for a production environment? I wouldn't trust this as-is — between the inconsistent classification, the fake-looking person count, and the lack of real-world validation from other users, this feels more like a research prototype or a demo than something you'd deploy in an actual home or office. If you just need basic presence, you'd be better off with a PIR sensor or mmWave radar that actually works out of the box. — Reply to this email directly, view it on GitHub <#252 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAWMM6RJJORWXBIACVCYYX3425V4FAVCNFSM6AAAAACWQRURBCVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTMOJTGQ3TIMQ> . You are receiving this because you commented.Message ID: ***@***.***>