ipro_reef_teach¶

Reef 的 Edge AI 韌體：一個映像包含三個 Teach 工具，image（USB 相機＋影像分類）、voice（聲紋辨識）與 sensor（感測器分類），開機時依 Flash 中儲存的模式擇一執行，由 IPRO AI Studio 操作。完整操作見 IPRO AI Teach

在 Reef 上執行¶

這個韌體使用自己的分割表 partition_reef_teach.toml（韌體 640 KB、mfg 1364 KB），第一次燒錄時連同 mfg 資料（開機模式與聲紋模型）一起寫入。影像與感測器模型之後由 IPRO AI Studio 經 USB 部署。

每個模式的 USB-C 識別碼不同（VID 0x3896）：image 為 PID 0x0705（UVC 相機＋序列埠），voice 為 0x0720、sensor 為 0x0721（序列埠）。不安裝 IPRO AI Studio 時，可以用瀏覽器版影像 Teach（python3 tools/teach_web/serve.py，以 Chrome 或 Edge 開啟 http://localhost:8780）。各模式的操作見 IPRO AI Teach。

README¶

One image for the IPRO7AI Reef that runs any of the three Teach tools, one at a time:

This is the one Edge AI firmware for the Reef; the source folders above hold each tool's code and are built only as part of it. The mode stored in flash picks which tool starts at boot, each with its own USB identity, so each IPRO AI Studio page finds the board. Switching stores another mode and restarts into it (about 2 s). Each tool keeps its model in its own slot, so switching loses none of them: deploy all three once and move between them freely.

Switching¶

• IPRO AI Studio: open the page for the tool you want. When the Reef is in another mode, the page says so and offers Switch to ….
• Console (UART0, 115200): mode shows the current one; mode image, mode voice or mode sensor restarts into another.
• USB: APP / APP <mode> on the voice and sensor serial protocols (answers APP <current> image voice sensor / APP OK <mode>, then the board restarts); on the image one, a 16-byte header "TAPP", u32 mode (0 image, 1 voice, 2 sensor, 0xFFFFFFFF to ask), u32 0, u32 0 (answers TAPP <current> image voice sensor / TAPP OK <mode>).

Build and flash¶

make -C apps/ai/ipro_reef_teach

This firmware needs its own partition table, partition_reef_teach.toml: the firmware partition is cut to 640 KB (the image is about 360 KB) and mfg grows to 1364 KB so all three models fit at once. PSM, backup, media and factory stay where the stock table has them.

The voiceprint model is flashed with the firmware, together with the mode to start in; the image and Sensor models are deployed from IPRO AI Studio later. Build the mfg payload, then flash everything:

python3 apps/ai/ipro_reef_teach/mfg_payload.py --voice apps/ai/ipro_reef_teach/model/model_b.spkb --mode image \
    -o reef_teach_mfg.bin
TOOL=<IntelProCube dir>/ipro_iot_tool_macos
$TOOL --chipname ipro7 --chip-variant IPRO7AI --port <download UART> --baudrate 921600 \
    --skip-ota --pt apps/ai/ipro_reef_teach/partition_reef_teach.toml \
    --firmware apps/ai/ipro_reef_teach/build/build_out/ipro_reef_teach_IPRO7.bin \
    --mfg reef_teach_mfg.bin

The tool exits 0 even when it fails; look for [All Success]. Leave out --mfg to update only the firmware: the mode and the three models stay.

At boot the console shows the mode:

RTCH:   mode: voice (`mode <image|voice|sensor>` to switch)
SPKM: model 'b3-e12': 437200 B, trunk 16/16 layers on the NPU, ...
RTCH: voice ready - shell `spk`, IPRO AI Studio Audio page over USB-C

The mfg partition¶

src/reef_mode.h has the map and CMakeLists.txt passes it to the three tools' sources (VOICE_MFG_*, TEACH_MFG_*, SENSOR_MFG_*).

Each slot ends with a 4 KB record of the model store (components/ai/store): a deploy is committed only after it has been written and read back, so a reset during one leaves the slot's model either complete or reported missing (install interrupted), never half written. A slot's room is its size less that sector: voice 444 KB, image 508 KB, Sensor 252 KB. mfg_payload.py writes the voice model's record too, so a model flashed with --mfg is found CRC-checked (SPKM: store: install #1, CRC … checked). Models deployed by the V1.1 firmware are still found (V1.1 payload, no record).

How it is put together¶

• main.c reads the mode and starts one tool (src/reef_apps.c) from a task at that tool's own start-up priority.
• cherryusb calls usbd_event_handler and usbd_cdc_acm_set_dtr by name. Built here (REEF_TEACH_COMBINED), each tool's are renamed <mode>_usbd_... and src/reef_usb.c forwards to the running one.
• .config is the union of the three tools' configurations, without CONFIG_SINGLE_PRECISION_FLOAT: the Sensor front end has to equal IPRO AI Studio's value for value. The camera sources get -fsingle-precision-constant on their own, as in their firmware.
• All three tools' static data is linked in (207 KB, in PSRAM); the large buffers (camera frames, sample rings, deploy staging) are allocated at start by the running tool only.
• Every tool answers the IPRO Edge AI platform's V2.0 queries (device protocol 2 - IPRO Edge AI Programming Guide, Device protocol chapter) with fw=ipro_reef_teach; CAPS/TCAP lists the three modes. IPRO AI Studio's ipro-ai model list --all walks through them and comes back.
• The AON watchdog is armed (CONFIG_APP_WATCHDOG, 8 s nominal, about 5.6 s on the Reef's RC clock) and fed from a top-priority task by hardware-timer time (components/os/freertos/app_watchdog.h): a stalled SoC restarts in the mode it was in, with its models. Console: watchdog, watchdog test.
• Test commands on the UART shell: multi [runs] (voice mode: the Sensor model opened beside the voice trunk and both run from two tasks, every output checked against the model alone), sens poll <Hz> <s> and sens prof (sensor mode).

Measured (2026-10-31, Reef)¶

The image mode's camera pipeline was not run in this test: the test Reef has no camera fitted.

The tools, their pages, the console commands and the device protocol are described in the IPRO7AI Reef user guide and the IPRO Edge AI Programming Guide.