AI Agent Applications

Overall block diagram

QQ / Channel
    ↓
OpenClaw Gateway (port 18789)
    ↓
rkllm3-server (port 7878) ──── 8-core NPU ──── Qwen3-1.7B
                              (mutually exclusive with VLM/OCR)

If you only want to use OpenClaw, jump to OpenClaw on RK182X.

1. Which AI Agent Frameworks Can Run on RK182X?

OpenClaw is verified running on the board.

Verify the OpenClaw process:

ps -ef | grep openclaw | grep -v grep

Output:

linaro      2149    2124  1 14:28 ?        00:00:41 openclaw-gateway

rkllm3-server shares the NPU with the VLM/OCR services and they cannot run at the same time: to run the LLM inference service, stop VLM/OCR first, then start it per rkllm3-server Startup Configuration (after nohup startup its ppid is 1, and the PID changes with each launch).

Verify the rkllm3-server process:

ps -ef | grep rkllm3 | grep -v grep

2. OpenClaw on RK182X

OpenClaw is a full-featured AI Agent platform that calls the OpenAI-compatible API provided by rkllm3-server.

2.1 Actual Version and Configuration

OpenClaw version:

openclaw --version

Output:

OpenClaw 2026.7.1-2 (0790d9f)

Command path:

which openclaw

Output:

/home/linaro/.nvm/versions/node/v25.9.0/bin/openclaw

Service status:

openclaw gateway status 2>&1 | head -10

Output:

Service: systemd user (enabled)
File logs: /tmp/openclaw/openclaw-2026-08-20.log
Command: /home/linaro/.nvm/versions/node/v25.9.0/bin/node /home/linaro/.nvm/versions/node/v25.9.0/lib/node_modules/openclaw/dist/index.js gateway --port 18789
Service file: ~/.config/systemd/user/openclaw-gateway.service
Service env: OPENCLAW_GATEWAY_PORT=18789

OpenClaw currently has 3 configuration warnings (version-manager PATH / Node, etc.); they do not affect operation but fixing them is recommended:

• PATH includes version managers
• Node from version manager
• System Node 22 LTS not installed

2.2 Configuration Files

~/.openclaw/openclaw.json:

cat ~/.openclaw/openclaw.json | head -30

Output:

{
  "meta": {
    "lastTouchedVersion": "2026.3.24",
    "lastTouchedAt": "2026-08-06T09:51:02.877Z"
  },
  "wizard": {
    "lastRunAt": "2026-04-08T02:37:53.595Z",
    "lastRunVersion": "2026.3.24",
    "lastRunCommand": "onboard",
    "lastRunMode": "local"
  },
  "acp": {
    "defaultAgent": "main"
  },
  "models": {
    "providers": {
      "minimax": {
        "baseUrl": "https://api.minimaxi.com/anthropic",
        "apiKey": "sk-*** (redacted)",
        "api": "openai-completions",
        "models": [
          {
            "id": "MiniMax-Text-01",
            "name": "MiniMax-Text-01",
            ...

cat ~/.openclaw/openclaw.json | grep -A 5 '"minimax"'

Output:

"minimax": {
  "baseUrl": "https://api.minimaxi.com/anthropic",
  "apiKey": "sk-*** (redacted)",
  "api": "openai-completions",
  ...
}

The board currently has only one provider, MiniMax. The actual provider is whatever setup.sh configures (may be minimax / deepseek / glm, etc.); see ch05 One-Click Deployment for details.

2.3 Starting the Services

# How it is actually started on the board
openclaw gateway

# rkllm3-server
rkllm3-server \
  -m /userdata/models/Qwen3-1.7B/Qwen3-1.7B.rknn \
  --weight /userdata/models/Qwen3-1.7B/Qwen3-1.7B.weight \
  --vocab /userdata/models/Qwen3-1.7B/Qwen3-1.7B.tokenizer.gguf \
  --embed /userdata/models/Qwen3-1.7B/Qwen3-1.7B.embed.bin \
  --embed-mmap \
  -a Qwen3-1.7B \
  --host 127.0.0.1 \
  --port 7878 \
  -c 0xff \
  -n 512

The --weight argument is mandatory; startup fails without it.

2.4 Verifying the Connection

Start rkllm3-server first (mutually exclusive with VLM/OCR). Output content varies randomly with inference.

curl -s http://127.0.0.1:7878/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"Qwen3-1.7B","messages":[{"role":"user","content":"测试"}],"max_tokens":15}'

Output:

{
  "choices": [
    {
      "finish_reason": "length",
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "<think>\n好的，我现在需要处理用户提供的查询：\"测试\"。首先，"
      }
    }
  ],
  "created": 1787213136,
  "model": "Qwen3-1.7B",
  "system_fingerprint": "b5166-24345353",
  "object": "chat.completion",
  "usage": { "completion_tokens": 15, "prompt_tokens": 9, "total_tokens": 24 },
  "id": "chatcmpl-vpO1Ka18Fv59AZz3DnDDV6AbmSpK5TfS",
  "timings": {
    "prompt_n": 9,
    "prompt_ms": 94.885,
    "prompt_per_token_ms": 10.542,
    "predicted_n": 15,
    "predicted_ms": 118.728,
    "predicted_per_token_ms": 7.9152,
    "predicted_per_second": 126.339
  }
}

• Inference latency: prompt 95ms + predicted 119ms
• Token speed: ~95 tokens/s (prompt), 126 tokens/s (predicted)

3. rkllm3-server Startup Configuration

rkllm3-server is an OpenAI-compatible LLM inference backend.

3.1 Starting (on-board Qwen3-1.7B)

rkllm3-server \
  -m /userdata/models/Qwen3-1.7B/Qwen3-1.7B.rknn \
  --weight /userdata/models/Qwen3-1.7B/Qwen3-1.7B.weight \
  --vocab /userdata/models/Qwen3-1.7B/Qwen3-1.7B.tokenizer.gguf \
  --embed /userdata/models/Qwen3-1.7B/Qwen3-1.7B.embed.bin \
  --embed-mmap \
  -a Qwen3-1.7B \
  --host 127.0.0.1 \
  --port 7878 \
  -c 0xff \
  -n 512

rkllm3-server shares the NPU with the VLM/OCR services; stop VLM/OCR before starting it. When the service is not running, nothing listens on port 7878.

3.2 Verifying the Service

curl -s http://127.0.0.1:7878/v1/models

Output:

{
  "object": "list",
  "data": [
    {
      "id": "Qwen3-1.7B",
      "object": "model",
      "owned_by": "rknn",
      "meta": {
        "vocab_type": 2,
        "n_vocab": 151936,
        "n_ctx_train": -1,
        "n_embd": 2048,
        "n_params": 0,
        "size": 0
      }
    }
  ]
}

3.3 Sampling Parameters

rkllm3-server --help 2>&1 | grep -E "(temp|top-k|top-p|repeat|penalty)"

Output:

--temp N                                temperature (default: 0.8)
--top-k N                               top-k sampling (default: 40, 0 = disabled)
--top-p N                               top-p sampling (default: 0.9, 1.0 = disabled)
--repeat-penalty N                      penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled)
--presence-penalty N                    repeat alpha presence penalty (default: 0.0, 0.0 = disabled)

Parameter	Default	Purpose
--temp	0.8	Temperature
--top-k	40	top-k
--top-p	0.9	top-p
--repeat-penalty	1.0	Repeat penalty
--presence-penalty	0.0	Presence penalty
--frequency-penalty	0.0	Frequency penalty

4. Starting VLM Multimodal

ls -lh /userdata/models/Qwen3-VL-2B/

Output (excerpt):

总计 1.9G
-rwxr-xr-x 1 root   root    8.1K  8月10日 14:34 inspect.sh
-rw-r--r-- 1 linaro linaro   20M  8月17日 13:51 llm_Qwen3-VL-2B.rknn
-rw-r--r-- 1 linaro linaro 1.1G  8月17日 13:52 llm_Qwen3-VL-2B.weight
-rw-r--r-- 1 linaro linaro 3.0K  8月11日 22:53 OCRBench.py
-rw-r----- 1 root   root    11K  8月17日 15:24 ocr_server.py
-rwxr-xr-x 1 root   root   1.1K  8月10日 14:34 ocr.sh
drwxr-xr-x 4 linaro linaro 4.0K  8月17日 15:22 PaddleOCR-VL
-rw-r--r-- 1 linaro linaro 594M  8月17日 13:55 Qwen3-VL-2B.embed.bin
-rwxr-xr-x 1 linaro linaro 240K  8月20日 14:00 test.jpg
drwxr-xr-x 2 linaro linaro 4.0K  8月 5日 17:04 tokenizer
-rw-r--r-- 1 linaro linaro 4.2M  8月17日 13:50 vision_Qwen3-VL-2B.rknn
-rw-r--r-- 1 linaro linaro 229M  8月17日 13:54 vision_Qwen3-VL-2B.weight
-rw-r----- 1 root   root    11K  8月17日 14:01 vlm_server.py
-rwxr-xr-x 1 root   root    913  8月10日 14:34 vl.sh

Start the VLM multimodal service:

rkllm3-server \
  -m /userdata/models/Qwen3-VL-2B/llm_Qwen3-VL-2B.rknn \
  --weight /userdata/models/Qwen3-VL-2B/llm_Qwen3-VL-2B.weight \
  --model2 /userdata/models/Qwen3-VL-2B/vision_Qwen3-VL-2B.rknn \
  --weight2 /userdata/models/Qwen3-VL-2B/vision_Qwen3-VL-2B.weight \
  --vocab /userdata/models/Qwen3-VL-2B/Qwen3-VL-2B.tokenizer.gguf \
  --embed /userdata/models/Qwen3-VL-2B/Qwen3-VL-2B.embed.bin \
  --host 127.0.0.1 \
  --port 7878 \
  --img-start "<|vision_start|>" \
  --img-end "<|vision_end|>" \
  --img-content "<|image_pad|>" \
  --img-width 384 --img-height 384 \
  -c 0xff

5. Comparison of the Three Agent Options

Scenario	Recommended	Field-tested status
Web UI + multiple Channels	OpenClaw	Deployed
Pure API inference	rkllm3-server	Deployed (started on demand; exclusive with VLM/OCR)

6. Prerequisites

systemctl is-active rknn3.service

Output:

active

7. Full Verification Checklist

7.1 RK182X NPU Service

systemctl is-active rknn3.service && echo "✅ NPU service running"

Output:

active
✅ NPU service running

lspci -d 1d87:182a

Output:

0004:41:00.0 Processing accelerators: Rockchip Electronics Co., Ltd Device 182a (rev 01)

ls -l /dev/pcie-rkep-*

Output:

crw------- 1 root root 10, 123  8月20日 16:56 /dev/pcie-rkep-0004:41:00.0

7.2 rkllm3-server

Start rkllm3-server first, otherwise the checks below produce no output.

ps aux | grep rkllm3-server | grep -v grep

Output:

linaro    111435  1.0  5.0 1062804 137636 ?      Ssl  16:05   0:02 rkllm3-server -m /userdata/models/Qwen3-1.7B/Qwen3-1.7B.rknn --weight ... --vocab ... --embed ... --embed-mmap -a Qwen3-1.7B --host 127.0.0.1 --port 7878 -c 0xff -n 512

curl -s http://127.0.0.1:7878/v1/models

Output (while the service is running):

{
  "object": "list",
  "data": [
    {
      "id": "Qwen3-1.7B",
      "object": "model",
      "owned_by": "rknn",
      "meta": { ... }
    }
  ]
}

7.3 OpenClaw

ps aux | grep openclaw-gateway | grep -v grep

Output:

linaro      2149  1.0  5.3 18552316 430088 ?     Ssl  14:28   0:42 openclaw-gateway

curl -s http://127.0.0.1:18789/health

Output:

{ "ok": true, "status": "live" }

7.4 End-to-End Test

Start rkllm3-server first (mutually exclusive with VLM/OCR). Inference output is random.

curl -s http://127.0.0.1:7878/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"Qwen3-1.7B","messages":[{"role":"user","content":"测试"}],"max_tokens":15}' \
  | python3 -c "import sys, json; d=json.load(sys.stdin); print(d['choices'][0]['message']['content'])"

Output (measured while the service is running; content varies randomly):

<think>
好的，我现在需要处理用户提供的查询："测试"。首先，

8. Notes

Startup order: all components start in the order rknn3.service → rkllm3-server → openclaw-gateway > Port occupancy check: netstat -tulpn | grep -E '(7878|18789)' > Logs: journalctl -u rknn3.service -f and tail -f /tmp/openclaw/openclaw-*.log

9. FAQ

Symptom	Cause	Fix
openclaw gateway fails to start	Node version / PATH issues	openclaw doctor --repair
rkllm3-server reports core_mask 8192 ...	-c is not the context	LLM uses -c 0xff
port 7878 already in use	rkllm3-server / VLM started twice	Find the occupier with ss -tlnp | grep 7878
curl /v1/models gets no response	rkllm3-server not started	Run in background nohup rkllm3-server ... &
OpenClaw reports PATH/Node warnings	Node installed via nvm	openclaw doctor --repair
VLM and LLM services are mutually exclusive	They share NPU memory	Run only one; switch per scenario

10. Next Steps

• SKILL User Manual — invoke specific Skills
• LLM Inference — CLI / HTTP / Web, three ways
• Industrial Anomaly Detection — VLM in practice