Workbench sits in the middle layer of the ShiMetaPi suite — connected upward to the AIHub cloud training platform and downward to the Pico-G1 board. It does not train models itself, nor does it perform NPU inference directly; it is an integration-and-orchestration IDE.
┌────────────────────────────────────────────────────────────────────────────────┐
│ │
│ AIHub cloud Workbench IDE Pico-G1 board │
│ (aihub.shimetapi.cn) (Windows) (192.168.49.10) │
│ │
│ ┌──────────┐ ┌────────────────┐ ┌──────────┐ │
│ │ Training │ ───HTTPS──────▶ │ Models Tab │ ──SSH───▶ │ /mnt/ │ │
│ │ Repo │ │ one-click │ │ models/ │ │
│ │ Market │ │ deploy │ │ xmm/ │ │
│ └──────────┘ │ │ └──────────┘ │
│ │ Vision Tab │ ──HTTP──▶ shimeta-svp │
│ │ shimeta-svp │ shimeta-xmm │
│ │ shimeta-xmm │ │
│ │ │ ──HTTP──▶ shimeta-py │
│ │ shimeta-py IDE │ v1.0.4 │
│ │ │ (Python runtime) │
│ │ Agent Tab │ ──SSH───▶ shimetad │
│ │ Claude Code │ (passwordless) │
│ └────────────────┘ │
│ │
└────────────────────────────────────────────────────────────────────────────────┘
| Responsibility | Does NOT |
|---|
| Upload/manage training datasets | Does not control the board |
| Launch GPU training jobs | Does not send inference commands |
| Maintain the official model library and community market | Does not parse board inference results |
Export .xmm.zip model packages | Does not store board logs |
| Responsibility | Does NOT |
|---|
| AIHub account login, token caching | Does not join model training |
| Board SSH connections, device list management | Does not write NPU operators |
| Models Tab browsing, filtering, model details | Does not recompile models |
One-click deploy (push zip to the board's /mnt/models/xmm/) | Does not store bulk training data |
| shimeta-py IDE: edit Python scripts and run them on the board | Does not render YUV images |
| Vision Tab launches shimeta-svp / shimeta-xmm | Does not implement inference algorithms |
| Embedded "Object view" (物象) preview panel on the right (camera feed) | Does not store raw camera streams |
| Agent Tab: Claude Code natural-language operations | — |
| Right-side resource monitoring (CPU/MEM/TMP/IPC) | — |
| Tool | Responsibility | Invoked via |
|---|
| shimeta-py v1.0.4 | Python runtime (code edited in the IDE runs directly) | IDE over SSH |
| shimeta-svp | 9 precompiled scenario detections (person/face/vehicle/non-motorized vehicle/facial expression/driver monitoring/head count/fire & smoke/body keypoints) | Vision Tab button |
| shimeta-xmm | Inference of user-defined .xmm models | Vision Tab panel |
| Does NOT | Explanation |
|---|
| No cloud training | Model training happens on AIHub |
| No user login sessions | The account system lives in the IDE |
| No model repository | Models come from AIHub |
User finishes training on AIHub and exports .xmm.zip
│
▼
User picks a model in the IDE Models Tab → Model Market / User models
│
▼ clicks "Deploy to board"
│
IDE automatically pushes over SSH:
scp <model>.xmm.zip → root@192.168.49.10:/mnt/models/xmm/
│
▼
Board extracts to /mnt/models/xmm/<model>.xmm/
│
▼
Vision Tab → the model appears automatically in the shimeta-xmm panel
│
▼
User clicks "Run" → shimeta-xmm starts
│
▼
The right-side "Object view" panel shows the camera feed (with detection results)
User types in the IDE Terminal Tab
ssh root@192.168.49.10
shimeta-svp person
│
▼ SSH channel
Board executes the command; stdout returns to the IDE terminal
[on-board shimeta-camera or shimeta-xmm service]
├─ MJPEG stream (port 8080 /stream) ──HTTP──▶ [IDE right-side "Object view" panel]
├─ Detection result JSON (/api/status) ─HTTP──▶ [status panel]
└─ RTSP stream (port 554) ──RTSP──▶ [external playback in VLC]
User types a natural-language instruction in the IDE Agent Tab
"check the board's current IP and CPU usage"
│
▼ Claude Code parses it
│
Agent generates the SSH commands automatically and executes them
│
▼ SSH channel
Board-side shimetad executes; results return to the Agent
│
▼
Agent replies to the user in natural language
| Workbench hard dependencies | Workbench does NOT depend on |
|---|
| AIHub account (for login and model training links) | Board internet access (SSH only is enough) |
| Pico-G1 board (SSH port 22 reachable) | AIHub training history (only finished models are pulled) |
| Board-side shimetad or root account | mirrors repository (board tools ship with the firmware) |
| Board-side shimeta-py / shimeta-svp / shimeta-xmm installed | Board port 554 (optional, only for RTSP playback) |
| Credential | Storage | Notes |
|---|
| AIHub account | Workbench local (encrypted cache) | Only used for login and model training links |
| Board shimetad | No password | IDE connects over SSH automatically, no setup |
| Board root | Workbench device management (manual entry) | Default 123456; changeable in settings |
