Home / Guides / ESP32-S3 camera robot
How do you build a robot around an ESP32-S3 camera board?
An ESP32-S3 camera robot puts the camera, the microphones, the speaker and the brain on one small board in its head. The Mini v2 uses the Waveshare ESP32-S3-CAM-OV5640 board ($37.99 CAD on amazon.ca, seen 2026-11-02): an OV5640 camera, two microphones, an audio chip and amplifier, a 2 W speaker in the box, a battery charger and an 18-pin screen port, with 8 MB of PSRAM and 16 MB of flash. On the chip itself, finding a face runs at about 23 frames a second and picking out general objects at about 1 a second, so the board can aim a head at you but should hand "what is on the desk" to a computer.
What is on the board
- Chip: ESP32-S3, S3R8 module, 8 MB PSRAM, 16 MB flash.
- Camera: OV5640 on a 4.1 mm lens at f/2.8; the listing does not mention autofocus.
- Sound: two microphones, an ES8311 audio codec, an NS4150B amplifier and a 2 W speaker in the box.
- Other: a battery charger, a memory-card slot and an 18-pin SPI screen port (no screen on the board).
From the Waveshare product page and listing, read by 3D3D's research on 2026-11-02 (research note R5, section 3). The board's own size is not on the listing; the head is fitted to it by the checks.
What the chip can do
| Job | Time | Per second |
|---|---|---|
| Find a face (two-stage, 120 x 160) | about 43 ms | 23 |
| Find a face (one-stage, 224 x 224) | 137 ms | 7 |
| Recognise whose face (112 x 112, after finding it) | 256 ms | about 4 |
| General objects (YOLO11n, 320 x 320) | 992 ms | 1 |
Source: Espressif's esp-dl model readmes (human_face_detect, human_face_recognition, coco_detect), seen 2026-11-02, int8 models; the "per second" column is 1000 divided by the time. A 480 x 320 JPEG stream at 20 frames a second needs 600 to 1,000 KB/s, which the S3 meets (Espressif camera notes). Wake-word listening (ESP-SR) runs on the chip too.
In plain words: finding a face and turning the head to it is comfortable; telling a mug from a dog is not, so that job goes to a computer or the cloud.
How it is built in
The head is printed in five pieces: the shell, the face plate (bezel), a clear visor over the screen, a camera lens ring and a light seam. The board, the 2 inch screen and the speaker sit inside. On the rolling bodies two M3 screws go up through the body's lid into two heat-set inserts in the head; on the desk bodies the head hangs on the neck's servo horn. The brain board starts at least 11 mm above the head's seat so the inserts' bosses stay clear. Every head and body meets one interface, which is why all 40 pairs check out.
Its face, voice and moves run in the test room today. The program that runs them on the real ESP32-S3 chip is not built yet.
Steps
- Pick a body and a head. Every head carries the board, the screen and the speaker on the same seat, so any head fits any body.
- Check the fit. The board starts 11 mm above the head seat so the two M3 inserts and the cable hole stay clear; the checks prove it.
- Order the parts or ask 3D3D. The parts list links each part on amazon.ca; 3D3D can buy them, print the pieces, build and test.
- Print the pieces. PETG on a 0.4 mm nozzle; the clear visor, lens ring and light seam are clear PETG.
- Test it in the browser. The test room runs the face, voice and moves on your computer first.
Where to go next
All prices are in Canadian dollars before tax and shipping. Bought parts are priced from amazon.ca listings (seen 2026-11-04), plus 15% when 3D3D buys them. Printing, the build and test fee and finishing are 3D3D's own rates, listed on the prices page.