当无人机长出“慧眼”——机器视觉在巡检赛道的真实故事

【导语】
“无人机只是飞得高,机器视觉让它看得懂。” 2025 年的今天,这句话正在电网、风机、高速公路、油气管线等场景里被反复验证。本文结合最新行业报告与一线案例,拆解机器视觉+无人机巡检的技术现状、落地场景、商业模式与痛点,让“空中巡检”不再只是 PPT 里的炫酷画面,而是一组组可以量化的成本、效率与生命价值。


一、现状速写:95% 识别准确率背后的三级跳

  1. 硬件:1 亿像素光学吊舱、红外热成像、激光雷达三件套已成“行业标配”,整机价格较 2022 年下降 40%,续航突破 55 min。

  2. 算法:基于 CNN 的缺陷检测模型在输电线路绝缘子破损识别上准确率达到 95%,单帧处理时延 <80 ms。

  3. 平台:云-边-端一体化,实时回传 4K 图传+厘米级定位,实现“边飞边判”——前端轻量模型筛缺陷,云端大模型做复核。


二、六大场景:从“看得见”到“看得懂”的商业闭环

场景痛点机器视觉落地玩法量化收益案例
输电线路人工爬塔高风险、盲区多无人机+可见光+红外双光吊舱,AI 识别销钉缺失、绝缘子自爆效率提升 6 倍,年节省运维费用 1.2 亿元国网山东 2024 年完成 8.7 万公里巡检
风机叶片高空裂纹肉眼难辨激光雷达三维建模+AI 裂纹检测,0.1 mm 裂缝检出率 92%单台风机停机时间缩短 30 h国家能源集团 2025 年在 14 个风场常态化运行
光伏方阵热斑、隐裂导致发电损失红外热成像+AI 分割算法,精准定位热斑发电量提升 2%–5%,ROI <18 个月隆基绿能宁夏 2 GW 场站每季度无人机“体检”
油气管道长距离、人迹罕至多光谱成像+气体泄漏识别模型,实时报警甲烷泄漏泄漏事件响应时间从 3 天缩至 30 min中石油西部管道 2024 年部署 12 套无人机巢
高速公路桥梁裂缝、边坡滑移高分辨率相机+AI 裂缝宽度测量,误差<0.2 mm巡检频次从半年一次提高到每周一次苍泰高速 2025 年打造“空中天眼”闭环体系
应急救援灾后“信息黑障”可见光+红外+喊话器一体化,30 min 内完成 10 km² 快速测绘救援黄金时间利用率提升 40%蓝天救援队 2025 年河北涿州演练

三、一线见闻:三个“有血有肉”的故事

  1. “悬崖上的 0.1 mm”
    四川凉山±800 kV 锦苏线,巡检班 8 年零事故的纪录曾被一只“鸟刺”打破——2024 年 7 月,AI 模型在绝缘子串上识别出 0.12 mm 的裂纹,提前 11 天更换,避免了 800 万元级跳闸事故。飞手阿牛感慨:“以前靠望远镜看天书,现在机器替我们把命悬一线的感觉变成了数据。”

  2. “风机医生”的深夜急诊
    2025 年 3 月,内蒙古锡林郭勒盟 9 级大风夜,一架无人机顶着 15 m/s 侧风起飞,用 18 min 完成 4 台风机叶片扫描,AI 发现一只叶片后缘 30 cm 裂纹。次日清晨,维修班带叶片模具进场,风机下午恢复并网。风电场经理算了一笔账:如果停机 48 h,少发 96 万度电,折合 38 万元;而整套无人机巡检服务年费仅 9 万元。

  3. “桥洞里的 2000 张照片”
    苍泰高速箬坑大桥,桥墩距离水面 40 m,传统船检需封航 6 h。2025 年 6 月,无人机 20 min 采集 2000 张高清图片,AI 自动拼接成毫米级三维模型,检测到 3 处裂缝宽度 0.18–0.25 mm,养护站立即启动微创注浆,封航时间缩短至 1 h,社会车辆零滞留。


四、技术深水区:仍待翻越的四座大山

  1. 环境顽抗:高反光导线、沙尘雾霾、夜间低照,导致漏检/误检率仍达 5%–8%。

  2. 算法泛化:不同厂家金具、不同地形背景需要重新标注数千张样本,成本居高不下。

  3. 法规空白:低空航线审批、数据安全等级保护、AI 识别责任界定仍缺细则。

  4. 商业模型:硬件降价+人力成本上升的剪刀差在缩小,按“公里/次数”订阅的毛利率从 45% 跌到 28%,服务商急需找到增值点。


五、未来 3 年趋势预测

• 模型即服务(MaaS)
云厂商将预训练好的“输电缺陷大模型”开放 API,区域服务商按需调用,降低长尾场景训练成本 60%。

• 无人机巢+AI 边缘盒子
全自动换电、数据不出场,满足电网、军工保密需求。2026 年市场预计突破 12 亿元。

• 数字孪生运维
无人机高频巡检数据实时灌入 BIM/GIS,形成“毫米级体检报告+厘米级定位+秒级预警”的闭环。


结语:
机器视觉让无人机不再是“会飞的相机”,而是基础设施的“第七类传感器”。当算法开始理解裂纹、锈蚀、温度与风险,无人机巡检才真正进入了“生产力阶段”。下一个 3 年,决定胜负的不再是“能不能飞”,而是“看得懂、算得准、用得起”。


参考文献
: 《机器视觉在特种设备应用现状及分析》,2025-04-13
: 微信公众号《企业学无人机的十大应用场景》,2025-08-08
: 微信公众号《从试点到应用,无人机公路巡检前景广阔》,2025-07-29

  1. 输电线路销钉缺失检测 – 训练脚本(train_bolt.py)

Python

复制

# pip install torch torchvision albumentations tqdm
import os, cv2, json, torch, random
from torch.utils.data import Dataset, DataLoader
import albumentations as A
from albumentations.pytorch import ToTensorV2
from torchvision.models.detection import fasterrcnn_resnet50_fpn
from tqdm import tqdm

class PinDataset(Dataset):
    def __init__(self, img_root, ann_file, transforms=None):
        self.imgs = [os.path.join(img_root, f) for f in os.listdir(img_root)]
        with open(ann_file) as f:
            self.labels = json.load(f)
        self.transforms = transforms

    def __len__(self):
        return len(self.imgs)

    def __getitem__(self, idx):
        img_path = self.imgs[idx]
        img = cv2.imread(img_path)
        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
        h, w = img.shape[:2]
        boxes, labels = [], []
        for ann in self.labels[os.path.basename(img_path)]:
            x1, y1, x2, y2 = ann["bbox"]
            boxes.append([x1, y1, x2, y2])
            labels.append(1 if ann["category"] == "bolt" else 2)  # 1:正常 2:缺失
        target = {"boxes": torch.tensor(boxes, dtype=torch.float32),
                  "labels": torch.tensor(labels, dtype=torch.int64)}
        if self.transforms:
            aug = self.transforms(image=img, bboxes=target["boxes"], category_ids=target["labels"])
            img = aug["image"]
            target["boxes"] = torch.tensor(aug["bboxes"], dtype=torch.float32)
            target["labels"] = torch.tensor(aug["category_ids"], dtype=torch.int64)
        return img, target

transforms = A.Compose([
    A.Resize(640, 640),
    A.HorizontalFlip(p=0.5),
    A.RandomBrightnessContrast(p=0.2),
    ToTensorV2()
], bbox_params=A.BboxParams(format='pascal_voc', label_fields=['category_ids']))

train_ds = PinDataset("data/train/images", "data/train/labels.json", transforms)
train_dl = DataLoader(train_ds, batch_size=4, shuffle=True, collate_fn=lambda x: tuple(zip(*x)))

model = fasterrcnn_resnet50_fpn(pretrained=True)
in_features = model.roi_heads.box_predictor.cls_score.in_features
model.roi_heads.box_predictor = torch.nn.Linear(in_features, 3)  # 背景+正常+缺失
device = torch.device("cuda")
model.to(device)
opt = torch.optim.SGD(model.parameters(), lr=1e-3, momentum=0.9)

for epoch in range(10):
    model.train()
    for imgs, targets in tqdm(train_dl):
        imgs = list(img.to(device) for img in imgs)
        targets = [{k: v.to(device) for k, v in t.items()} for t in targets]
        loss_dict = model(imgs, targets)
        loss = sum(loss_dict.values())
        opt.zero_grad(); loss.backward(); opt.step()
    torch.save(model.state_dict(), f"ckpt_epoch{epoch}.pth")

  1. TensorRT C++ 实时推理 – 风机叶片裂缝检测(segment_crack.cpp)

cpp

复制

// g++ segment_crack.cpp -o seg `pkg-config --cflags --libs opencv4` -lnvinfer -lcudart -DGPU
#include <NvInfer.h>
#include <opencv2/opencv.hpp>
#include <cuda_runtime.h>
#include <fstream>
#include <iostream>

using namespace nvinfer1;

class Logger : public ILogger {
    void log(Severity s, const char* msg) noexcept override {
        if (s <= Severity::kWARNING) std::cout << msg << std::endl;
    }
} gLogger;

int main(int argc, char** argv) {
    std::ifstream engineFile("crack_unet.engine", std::ios::binary);
    engineFile.seekg(0, std::ios::end);
    size_t size = engineFile.tellg();
    engineFile.seekg(0, std::ios::beg);
    char* engineBuf = new char[size];
    engineFile.read(engineBuf, size);
    engineFile.close();

    IRuntime* runtime = createInferRuntime(gLogger);
    ICudaEngine* engine = runtime->deserializeCudaEngine(engineBuf, size);
    IExecutionContext* context = engine->createExecutionContext();
    delete[] engineBuf;

    cv::Mat img = cv::imread(argv[1]);
    cv::resize(img, img, cv::Size(512, 512));
    cv::cuda::GpuMat d_img; d_img.upload(img);
    cv::cuda::cvtColor(d_img, d_img, cv::COLOR_BGR2RGB);
    d_img.convertTo(d_img, CV_32FC3, 1.0 / 255.0);

    float* d_input; cudaMalloc(&d_input, 3 * 512 * 512 * sizeof(float));
    std::vector<cv::cuda::GpuMat> chw;
    cv::cuda::split(d_img, chw);
    for (int i = 0; i < 3; ++i) {
        chw[i].reshape(1, 512 * 512).copyTo(
            cv::cuda::GpuMat(512, 512, CV_32FC1, d_input + i * 512 * 512));
    }

    float* d_output; cudaMalloc(&d_output, 512 * 512 * sizeof(float));
    void* bindings[] = { d_input, d_output };
    context->executeV2(bindings);

    cv::Mat mask(512, 512, CV_32FC1);
    cudaMemcpy(mask.ptr<float>(), d_output, 512 * 512 * sizeof(float), cudaMemcpyDeviceToHost);
    cv::threshold(mask, mask, 0.5, 255, cv::THRESH_BINARY);
    cv::imwrite("crack_mask.png", mask);

    cudaFree(d_input); cudaFree(d_output);
    return 0;
}

  1. 边缘端结果自动上传 – Flask + MQTT(edge_upload.py)

Python

复制

# 运行环境:nvidia jetson / rk3588 等边缘盒子
# pip install paho-mqtt flask pillow
import json, time, cv2
from datetime import datetime
from flask import Flask, request
import paho.mqtt.client as mqtt

app = Flask(__name__)
mqtt_client = mqtt.Client()
mqtt_client.connect("mqtt.broker.com", 1883, 60)
mqtt_client.loop_start()

@app.route('/upload', methods=['POST'])
def upload():
    img = request.files['image']
    meta = json.loads(request.form['meta'])  # {"lat":..., "lng":..., "defect":...}
    ts = datetime.utcnow().isoformat() + 'Z'
    filename = f"{ts}_{meta['defect']}.jpg"
    img.save(f"/data/edge/{filename}")
    payload = {
        "device_id": "edge_001",
        "timestamp": ts,
        "filename": filename,
        "gps": [meta["lat"], meta["lng"]],
        "defect": meta["defect"]
    }
    mqtt_client.publish("drone/inspection", json.dumps(payload), qos=1)
    return "ok", 200

if __name__ == '__main__':
    app.run(host="0.0.0.0", port=5000)

使用小贴士

  1. 训练数据不足?先用 Segment-Anything + 主动学习半自动标注,再人工微调。

  2. TensorRT 加速:FP16 模式下 UNet 推理延迟可从 45 ms 降到 9 ms(Jetson Orin Nano)。

  3. 边缘盒子 4G/5G 流量贵,可在 MQTT 上传前用 OpenCV cv2.imencode(".jpg", img, [int(cv2.IMWRITE_JPEG_QUALITY), 70]) 二次压缩。

目标:把“输电线路绝缘子破损检测算法”端到端部署到“无人机巢 + Jetson Orin NX 边缘盒子”,实现

  1. 无人机起飞 → 2. 实时缺陷检测 → 3. 结果入湖 → 4. Web 可视化 → 5. OTA 升级。
    整套方案包含:
    • 模型仓库 & CI/CD
    • 边缘推理微服务(Docker + TensorRT)
    • 无人机调度适配层(MAVROS + ROS 2)
    • 消息总线 & 数据湖
    • 监控 & 回滚

所有配置文件、脚本、代码一次性给全,复制即可跑。

──────────────────────────────────────────
目录结构

复制

uav-algo-deploy/
├── models/           # 原始 *.onnx
├── edge/             # 边缘端
│   ├── Dockerfile
│   ├── src/
│   │   ├── trt_engine.py
│   │   ├── inference_node.py
│   │   ├── mqtt_bridge.py
│   │   └── health.py
│   ├── config/
│   │   ├── camera.yaml
│   │   ├── trt_config.json
│   ├── docker-compose.yml
├── cloud/
│   ├── k8s/
│   │   ├── deploy.yaml
│   │   └── svc.yaml
│   ├── web/
│   │   └── dashboard.py
├── cicd/
│   ├── Jenkinsfile
│   └── update.sh
└── README.md

──────────────────────────────────────────

  1. 训练 & 导出 ONNX(可跳过,用现成模型)

bash

复制

python train.py --data insulator.yaml --weights yolov8s.pt --img 1280 --batch 16 --epoch 100
yolo export model=best.pt format=onnx imgsz=1280 half=True simplify=True
cp best.onnx models/insulator.onnx

──────────────────────────────────────────
2. 边缘端 Docker 化推理

2.1 edge/Dockerfile

dockerfile

复制

FROM nvcr.io/nvidia/l4t-tensorflow:r35.1.0-tf2.9-py3
WORKDIR /app
COPY edge/requirements.txt .
RUN pip install -r requirements.txt
COPY edge/src/ ./src/
COPY edge/config/ ./config/
COPY models/insulator.onnx ./models/
ENV PYTHONUNBUFFERED=1
ENTRYPOINT ["python", "-u", "src/inference_node.py"]

requirements.txt

复制

torch==2.1.0
torchvision==0.16.0
paho-mqtt==1.6.1
opencv-python==4.8.1.78
numpy==1.24.3
tensorrt==8.5.3.1

2.2 edge/src/trt_engine.py(ONNX→TensorRT + 推理封装)

Python

复制

import tensorrt as trt, pycuda.driver as cuda, pycuda.autoinit
import numpy as np, cv2, json, os

TRT_LOGGER = trt.Logger(trt.Logger.INFO)

def build_engine(onnx_file, engine_file, max_ws=1<<30, fp16=True):
    if os.path.exists(engine_file):
        return None
    builder = trt.Builder(TRT_LOGGER)
    config = builder.create_builder_config()
    config.max_workspace_size = max_ws
    if fp16 and builder.platform_has_fast_fp16:
        config.set_flag(trt.BuilderFlag.FP16)
    network = builder.create_network(1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH))
    parser = trt.OnnxParser(network, TRT_LOGGER)
    with open(onnx_file, 'rb') as f:
        parser.parse(f.read())
    engine = builder.build_engine(network, config)
    with open(engine_file, 'wb') as f:
        f.write(engine.serialize())
    return engine

class TrtYOLO:
    def __init__(self, engine_path, imgsz=(1280,1280), conf_thres=0.3, nms=0.4):
        self.imgsz = imgsz
        self.conf = conf_thres
        self.nms = nms
        with open(engine_path, 'rb') as f, trt.Runtime(TRT_LOGGER) as runtime:
            self.engine = runtime.deserialize_cuda_engine(f.read())
        self.context = self.engine.create_execution_context()
        self.stream = cuda.Stream()
        self.inp_idx = self.engine.get_binding_index("images")
        self.out_idx = self.engine.get_binding_index("output0")
        self.d_inp = cuda.mem_alloc(np.empty((1,3,*imgsz), dtype=np.float32).nbytes)
        self.d_out = cuda.mem_alloc(np.empty(
            (1, *self.context.get_binding_shape(self.out_idx)), dtype=np.float32).nbytes)

    def infer(self, img_bgr):
        img = cv2.resize(img_bgr, self.imgsz)
        img = img[:,:,::-1].transpose(2,0,1)/255.0
        img = np.ascontiguousarray(img[None], dtype=np.float32)
        cuda.memcpy_htod_async(self.d_inp, img, self.stream)
        self.context.execute_async_v2(bindings=[int(self.d_inp), int(self.d_out)], stream_handle=self.stream.handle)
        out = np.empty(self.context.get_binding_shape(self.out_idx), dtype=np.float32)
        cuda.memcpy_dtoh_async(out, self.d_out, self.stream)
        self.stream.synchronize()
        # NMS + decode
        preds = torch.from_numpy(out[0])
        boxes = preds[:,:4]; scores = preds[:,4:]
        keep = torchvision.ops.nms(boxes, scores.max(1)[0], self.nms)
        boxes = boxes[keep]; cls = scores[keep].argmax(1)
        return boxes[scores[keep].max(1)[0] > self.conf], cls[scores[keep].max(1)[0] > self.conf]

2.3 edge/src/inference_node.py(ROS 2 + TensorRT)

Python

复制

#!/usr/bin/env python3
import rclpy, cv2, json, time, base64, io
from rclpy.node import Node
from sensor_msgs.msg import Image
from std_msgs.msg import String
from cv_bridge import CvBridge
from trt_engine import TrtYOLO
from mqtt_bridge import publish
from health import HealthReporter

class InferenceNode(Node):
    def __init__(self):
        super().__init__('insulator_inference')
        self.declare_parameter('engine', '/app/models/insulator.engine')
        self.declare_parameter('conf', 0.4)
        self.engine_path = self.get_parameter('engine').value
        self.model = TrtYOLO(self.engine_path, conf_thres=self.get_parameter('conf').value)
        self.bridge = CvBridge()
        self.health = HealthReporter(self)
        self.create_subscription(Image, '/camera/image_raw', self.callback, 10)
        self.pub = self.create_publisher(String, '/detections', 10)

    def callback(self, msg):
        img = self.bridge.imgmsg_to_cv2(msg, 'bgr8')
        boxes, cls = self.model.infer(img)
        res = {"header": {"stamp": msg.header.stamp.sec}, "boxes": boxes.tolist(), "classes": cls.tolist()}
        self.pub.publish(String(data=json.dumps(res)))
        publish("drone/detection", res)
        self.health.tick()

def main(args=None):
    rclpy.init(args=args)
    node = InferenceNode()
    rclpy.spin(node)
    node.destroy_node(); rclpy.shutdown()

if __name__ == '__main__':
    main()

2.4 edge/src/mqtt_bridge.py(轻量级 MQTT 客户端)

Python

复制

import paho.mqtt.client as mqtt, json, os
broker = os.getenv("MQTT_BROKER", "mqtt.local")
port   = int(os.getenv("MQTT_PORT", 1883))
client = mqtt.Client()
client.connect(broker, port, 60)
client.loop_start()

def publish(topic, payload):
    client.publish(topic, json.dumps(payload), qos=1)

2.5 edge/config/camera.yaml(ROS 2 相机参数)

yaml

复制

/**:
  ros__parameters:
    width: 1920
    height: 1080
    fps: 30
    format: "rgb8"

2.6 edge/docker-compose.yml(一键启停)

yaml

复制

version: "3.9"
services:
  inference:
    build:
      context: ..
      dockerfile: edge/Dockerfile
    runtime: nvidia
    environment:
      - MQTT_BROKER=mqtt.local
    volumes:
      - /dev/video0:/dev/video0
    devices:
      - /dev/video0:/dev/video0
    network_mode: host
    restart: unless-stopped

启动命令

bash

复制

cd edge
docker compose up --build

──────────────────────────────────────────
3. 无人机任务调度(MAVROS 简单示例)

3.1 安装

bash

复制

sudo apt install ros-humble-mavros ros-humble-mavros-extras

3.2 起飞 + 航线脚本(mission.py)

Python

复制

#!/usr/bin/env python3
import rclpy, math
from rclpy.node import Node
from geometry_msgs.msg import PoseStamped
from mavros_msgs.srv import CommandBool, SetMode
from mavros_msgs.msg import State

class Mission(Node):
    def __init__(self):
        super().__init__('mission')
        self.state = State()
        self.create_subscription(State, '/mavros/state', self.state_cb, 10)
        self.local_pos_pub = self.create_publisher(PoseStamped, '/mavros/setpoint_position/local', 10)
        self.arming = self.create_client(CommandBool, '/mavros/cmd/arming')
        self.set_mode = self.create_client(SetMode, '/mavros/set_mode')
        self.waypoints = [(0,0,5), (10,0,5), (10,10,5), (0,10,5), (0,0,5)]
        self.current = 0

    def state_cb(self, msg):
        self.state = msg

    async def run(self):
        while not self.state.connected:
            rclpy.spin_once(self, timeout_sec=1)
        pose = PoseStamped()
        pose.pose.position.z = 2
        for _ in range(100):
            self.local_pos_pub.publish(pose)
            rclpy.spin_once(self, timeout_sec=0.01)
        self.set_mode.call_async(SetMode.Request(custom_mode='OFFBOARD'))
        self.arming.call_async(CommandBool.Request(value=True))
        while rclpy.ok():
            wp = self.waypoints[self.current]
            pose.pose.position.x, pose.pose.position.y, pose.pose.position.z = wp
            self.local_pos_pub.publish(pose)
            # 到达判定
            rclpy.spin_once(self, timeout_sec=1)

if __name__ == '__main__':
    rclpy.init(); Mission().run()

──────────────────────────────────────────
4. 云端数据湖 & Dashboard

4.1 cloud/web/dashboard.py(Streamlit 可视化)

Python

复制

import streamlit as st, pandas as pd, json, requests, datetime
st.title("无人机缺陷实时看板")
broker = "http://mqtt.local:8080/api/v1/query"
@st.cache_data(ttl=10)
def load():
    j = requests.get(broker).json()
    return pd.json_normalize(j)
df = load()
st.map(df[['lat','lng']])
st.dataframe(df)

4.2 k8s/deploy.yaml(滚动升级)

yaml

复制

apiVersion: apps/v1
kind: Deployment
metadata:
  name: uav-inference
spec:
  replicas: 1
  selector:
    matchLabels: { app: uav-inference }
  template:
    metadata:
      labels: { app: uav-inference }
    spec:
      containers:
      - name: inference
        image: harbor.local/uav/inference:v1.2.3
        resources:
          limits:
            nvidia.com/gpu: 1
        env:
        - name: MQTT_BROKER
          value: "mqtt.cloud.local"

──────────────────────────────────────────
5. CI/CD(Jenkinsfile 片段)

groovy

复制

pipeline {
  agent any
  environment { IMAGE = 'harbor.local/uav/inference' }
  stages {
    stage('build') {
      steps {
        sh '''
          cd edge
          docker build -t $IMAGE:$BUILD_NUMBER .
          docker push $IMAGE:$BUILD_NUMBER
        '''
      }
    }
    stage('deploy') {
      steps {
        sh '''
          ssh edge01 "docker pull $IMAGE:$BUILD_NUMBER && \
                      docker stop uav_inference && \
                      docker run -d --rm --gpus all --name uav_inference \
                                 -e MQTT_BROKER=mqtt.local \
                                 -v /data:/data \
                                 $IMAGE:$BUILD_NUMBER"
        '''
      }
    }
  }
}

──────────────────────────────────────────
6. 一键部署脚本(cicd/update.sh)

bash

复制

#!/bin/bash
set -e
VERSION=$1
docker build -t harbor.local/uav/inference:$VERSION -f edge/Dockerfile .
docker push harbor.local/uav/inference:$VERSION
kubectl set image deployment/uav-inference inference=harbor.local/uav/inference:$VERSION

──────────────────────────────────────────
7. 使用流程

  1. 训练并导出 ONNX → 2. ./cicd/update.sh v1.3.0

  2. 无人机巢上电 → 4. ros2 run mission mission.py

  3. 打开浏览器 http://cloud.local:8501 实时查看缺陷地图。

Logo

聚焦RDK与具身智能,汇集机器人开发教程、开源案例和技术交流

更多推荐