人行道配送机器人如何在城市里找到方向？How Do Sidewalk Delivery Robots Find Their Way Around Cities?
传感器融合与导航是本文拆解的两大支柱：看看激光雷达和 AI 路径规划如何让自主配送机器人在 2026 年安全穿行错综复杂的城市人行道。
Sensor fusion, LiDAR and AI-driven route computation power the autonomous delivery bots that roll safely through complicated urban footways. Here is how the stack works in 2026.
不妨想象这样一台小巧的六轮小车，在人来人往的人行道上稳稳前行，从行人之间穿过去，给停着的电动滑板车让出一大块空间，到了路口又耐着性子等车流清空。它的货舱里装着你的午饭，或是附近小店送来的包裹，却能在一片混乱的街景里走得从容，看着毫不费力。这背后究竟是怎么运作的？
凡是好奇这些小车怎样在城市人行道上穿行的人，马上就能看到支撑人行道自主行驶的整套技术。别再把它当成遥控玩具：这是相当严谨的机器人平台，配齐了一整套传感器、人工智能和实时判断能力，某些地方甚至能与自动驾驶汽车一较高下。
如今的配送机器人一般都装有摄像头、激光雷达、用于探测障碍的超声波传感器，以及用于规划路线的 GPS 导航 [[11]]。而硬件清单还只是个开头。借助传感器融合，它们把 10-12 个摄像头、多台激光雷达、雷达和超声波设备的数据汇到一起，在周身形成一个完整的 360 度「感知气泡」[[23]]。有了这幅拼合而成的图景，机器便能在错综复杂的街道上安全行进，无论是乱穿马路的行人，还是停得歪七扭八的电动滑板车，都能远远绕开 [[1]]。
01传感器阵容：机器的眼睛与耳朵
要看懂人行道导航，得先看那套赋予机器人感知能力的精密传感器阵列。它并不依赖单一技术，而是让一组经过精心协调的设备协同工作，彼此弥补对方的短板。
一台标准配送机器人身上大约有 10 个摄像头、4 台雷达、8 个超声波传感器和若干台激光雷达 [[20]]。装在这么一台小车上听着奢侈，但每一类传感器都在拼合完整环境画面时承担着不可或缺的角色。
传感器融合为什么关键
真正令人赞叹的，是融合算法让所有这些设备协同运转的时刻。激光雷达测距精准，却不容易判断物体类别；摄像头极擅长辨认物体，却会在昏暗光线或恶劣天气下失灵；超声波传感器近距离可靠，分辨率却有限。把各路数据汇集到一起，机器人对环境的把握便远胜于任何单一传感器。
多台摄像头同步工作，使静止物体和移动行人等障碍都能被迅速发现并识别 [[39]]。这种冗余十分关键：一旦 GPS 信号在城市峡谷中丢失，机器人就能依靠摄像头的视觉里程计和激光雷达 SLAM 维持自身定位。
02激光雷达：用三维描绘世界
就人行道行驶而言，激光雷达（LiDAR，Light Detection and Ranging）完全称得上最重要的单个传感器。它发射激光脉冲，测出每一束脉冲折返所需的时间，再由这些计时数据生成细节极为丰富的 3D 点云。
对配送机器人来说，激光雷达同时承担着好几项任务：
- 识别人行道：它能精确定位人行道与车道之间的分界，哪怕边缘模糊不清或被遮挡，也能确保机器走在正确的路面上。
- 测量障碍距离：摄像头只能估算距离，激光雷达却给出确切数值，让机器人毫不含糊地知道行人或障碍究竟有多远。
- 三维建图：机器会在周围实时搭建一幅三维地图，不仅标出有什么，还标出它在立体空间中位于何处。
如今的配送机器人采用抗阳光激光雷达，即便在强烈日光下也能有效工作 [[35]]。这一点至关重要，因为人行道行驶主要发生在白天，太阳干扰本可能让传感器读数被淹没。
SLAM：同步定位与地图构建
激光雷达数据会送入 SLAM 算法，让机器一边绘制陌生环境的地图，一边持续追踪自己在地图中的位置。在高楼林立的城区，这一点不可或缺——信号被楼宇反射（即「城市峡谷」效应），单靠 GPS 并不够。
快要驶近一个熟悉的路口时，机器人能调出过去在那里行之有效的策略；而遇到前所未见的情形，则可能加倍谨慎，必要时还会通过远程操控系统请人介入。
03计算机视觉：教会机器人「看」
激光雷达回答的是物体在哪里，计算机视觉回答的则是那是什么。配送机器人依靠在海量图像上训练而成的精密神经网络，对周围事物进行实时分类。
视觉系统必须识别并归类以下对象：
| 物体类别 | 导航应对 | 优先级 |
|---|---|---|
| 行人 | 让行、保持安全距离、预测行进轨迹 | 关键 |
| 人行横道 | 停下、等候行人通过、确认路面清空 | 关键 |
| 交通信号灯 | 在路口遵守通行与禁止通行信号 | 高 |
| 静止障碍 | 绕行通过（消防栓、长椅、垃圾桶） | 中 |
| 动态障碍 | 预测动向并调整路线（电动滑板车、自行车、宠物） | 高 |
| 人行道边缘 | 与路缘和车道分界线保持安全间距 | 高 |
用于物体识别的深度学习
卷积神经网络（CNN）与基于 Transformer 的视觉模型，如今让配送机器人的物体识别精度达到了人类水平。这些模型在数百万张城市环境图像上训练过，学会了辨认从儿童自行车到施工护栏的各种东西。
摄像头系统所采用的 AI 训练方法与机器人抓取所用的颇为相近。因此这套系统不只是发现物体，还能读懂它的属性——坚硬还是柔软、静止还是移动、危险还是无害。
语义分割
仅有物体检测还不够。语义分割会把摄像头画面中的每一个像素都归类，使机器人能区分可通行地面（人行道铺面）与不可通行区域（草地、车道、水景），哪怕两者之间没有任何实体阻隔。
04路径规划：坐镇指挥的 AI
传感器给了机器环境画面，视觉告诉它各种东西是什么，接下来就要解决如何从甲地抵达乙地。这正是路径规划算法登场之处。
路径规划是机器通过计算找出一条从起点到目标地点、全程无碰撞路线的过程 [[30]]。对人行道机器人来说，距离最短并不是关键：路线必须是最安全、最高效、也最符合社会规范的那一条。
分层规划
如今的配送机器人采用层级式规划：
- 全局规划：借助 GPS 和预装地图，先算出从餐厅或商店到收货地址的大致路线，把街道布局、已知人行道网络和合法过街点都考虑进去。
- 局部规划：随着机器前行，眼前的几米（接下来的 5-10 米）会被不断重新规划，以避开障碍、行人和突发状况——每秒要重算几十次。
- 轨迹优化：机器规划的并非单纯的几何线条，而是一条轨迹，把自身的运动限制——加速度、转弯半径、速度上限——都纳入考虑，从而保证动作平顺稳定。
针对人行道配送机器人的稳健路径规划研究表明，车身更宽、速度更慢、行驶更保守的机器人，尤其需要稳健的方法 [[4]]。这也解释了常见的一幕：机器人以行人速度（3-5 mph）行进，并远远绕开障碍 [[12]]。
在一次次出行中积累经验
机器学习让更先进的规划系统随时间不断精进。正如 Coco Robotics 和 Avride 在一趟趟配送中学到的，一座不肯按规矩出牌的城市对导航者究竟意味着什么 [[3]]。某段人行道中午永远拥挤、某个路口 GPS 信号总是很差——这些记忆都会直接反映到行为调整上。
这正是基础模型机器人技术发挥作用的地方：预先训练好的导航模型能够在不同城市和环境间迁移，减少部署前所需的大量本地建图工作。
05实时动态避障
空无一人的人行道说明不了什么。真正的考验是城市人流混乱而不可预测的现实，而应对它需要出色的动态避障能力。
遇到障碍时，配送机器人会依次完成以下步骤：
- 通过传感器阵列发现它
- 判断它属于哪一类——人、车辆，还是静止物体
- 预测它片刻之后可能出现的位置
- 规划一条能避开它的替代路线
- 平稳、安全地执行这一动作
预测建模
对于行人这类移动目标，预测模型会预判其未来位置。若有人径直朝机器人走来，系统会预测对方的轨迹并相应调整路线——也许放慢速度，或稍稍向右偏一点，以便安全错身。
物体趋近机制让机器人与近旁障碍保持安全距离，轨迹记忆则生成最优绕行策略 [[32]]。实际上，机器人绝不只是对眼前的东西作出反应：它始终维持着整个场景的工作模型，提前想好几步。
社交导航
不撞上去只是底线而非终点——配送机器人还必须合乎社交地行进。具体而言包括：
- 向行人让行，哪怕从规则上讲它本拥有路权
- 跟随他人时保持合理间距
- 通过动作或指示灯把意图表达清楚
- 避免突然、生硬的动作吓到旁人
这些社交规则被直接写入路径规划算法，使机器人既不会撞上公众，也不会沦为惹人厌烦的麻烦或安全隐患。
应对边缘情况
再先进的 AI，也会遇到无法独自处理的场面。遇到真正模棱两可的情形——比如人行道被施工彻底封死——机器人可通过远程操控系统把情况转给远端操作员，由人来提供它所欠缺的判断 [[1]]。
06城市导航中的难点
城市人行道会抛出种种独特难题，把自主导航技术逼到极限，也解释了配送机器人的部署为何一直缓慢而谨慎。
| 挑战 | 影响 | 解决方案 |
|---|---|---|
| GPS 信号丢失 | 城市峡谷会阻断卫星信号 | 激光雷达 SLAM 加视觉里程计进行航位推算 |
| 光线不佳 | 夜色与阴影削弱摄像头效果 | 激光雷达加雷达——不依赖光线的传感器 |
| 恶劣天气 | 雨、雪、雾都会降低传感器表现 | 传感器冗余，配合针对天气调校的算法 |
| 拥挤的人行道 | 密集人流让导航变得异常复杂 | 预测建模，配合保守的限速 |
| 施工区域 | 临时出现的障碍，改道的路线 | 实时建图，并保留远程操控作为后手 |
| 无障碍通行 | 正如对机器人挑战的分析所指出的，要通过路缘坡、坡道和不平整路面 | 先进悬架，配合地形分类 |
安全层面的挑战
物理导航只是其中一条战线，网络威胁同样必须防范。正如掌握 AI 深度伪造检测能让人免受篡改媒体的侵害，自主机器人也必须确认自己的传感器数据没有被伪造或入侵——感知一旦被污染，就可能把它推向危险的错误。
07人行道导航的未来走向
支撑配送机器人穿行城市人行道的技术仍在飞速演进。几个新兴趋势有望让这些系统更强大、也更普及：
以基础模型驱动导航
企业正设法利用庞大且不断增长的视频库，为自主人行道导航预训练导航基础模型 [[27]]。这类模型在数百万小时的城市导航数据上训练而成，有望让机器人只需极少的补充训练，便能适应一座新城市。
V2X（车联万物，Vehicle-to-Everything）通信
未来的配送机器人或许能直接与交通基础设施（智能红绿灯、人行横道信号）以及其他自动驾驶车辆通信，由此形成协同式导航生态，比每台机器各自孤立决策更加安全高效。
更先进的仿真与数字孪生
一座新城市已不必完全在真实街道上摸索。部署之前，机器人如今可以先在照片级逼真的仿真环境——真实城市的数字孪生——中进行大量训练，熟悉当地的种种怪癖，比如市中心那个古怪路口、永远挤满人的农贸市场，而不必冒现实事故的风险。
正如在 AI 机器人领域领先的国家所显示的，大力投资智慧城市基础设施的国家，正在打造配送机器人得以蓬勃发展的环境：专用机器人车道、维护更好的人行道，以及一体化的交通管理系统。
群体智能
多台配送机器人在同一区域作业时，可以实时共享导航数据。只要有一台发现人行道被封或出现新障碍，就能立刻通知附近其他机器人，这种集体感知能提升整个车队的效率与安全。
08常见问题
配送机器人靠什么在城市人行道上认路？
人行道配送机器人都带哪些传感器？
配送机器人能在人行道上躲开障碍吗？
配送机器人的行驶速度是多少？
机器人迷路或卡住时会怎么办？
恶劣天气下配送机器人还能工作吗？
Imagine a compact six-wheeled cruiser rolling along a crowded footway, threading past people on foot, giving parked e-scooters a wide margin, and sitting patiently until a crossing clears. Inside its cargo bay sits your lunch or a parcel from a nearby shop, and it cuts through the disorder of the street with a calm that looks almost effortless. What is actually going on under the hood?
Anyone curious about the way these bots thread through city footways is about to meet the technology stack that makes self-driving on pavements possible. Forget the image of a remote-controlled toy: these are serious robotic platforms packing a full armoury of sensors, artificial intelligence and live judgement calls — enough, in places, to rival an autonomous car.
Cameras, LiDAR, ultrasonic obstacle detectors and GPS-based routing are standard fittings on today's delivery bots [[11]]. Yet the hardware list is only where the story starts. Sensor fusion merges the feeds of 10-12 cameras, several LiDAR units, radar and ultrasonic devices into one all-round 360-degree "bubble of awareness" [[23]]. Armed with that merged picture, the machine moves safely through tangled streets, giving a wide berth to everything from jaywalkers to sloppily parked e-scooters [[1]].
01The Sensor Line-Up: Eyes and Ears of the Machine
Making sense of pavement navigation starts with the elaborate sensor array that gives a bot its sense of perception. No one technology carries the load; instead a carefully coordinated set of devices works together, each covering for the limits of another.
Roughly 10 cameras, four radar units, eight ultrasonic devices and a number of LiDAR sets ride aboard a standard delivery bot [[20]]. For a compact vehicle it sounds lavish, yet every class of sensor does an indispensable job in stitching together a full view of the surroundings.
Why Sensor Fusion Carries Weight
The genuinely impressive part arrives once fusion algorithms set all these devices working in concert. LiDAR nails distance but cannot easily say what an object is; cameras identify things beautifully yet stumble in dim light or rough weather; ultrasonic units are dependable up close but coarse. Pool every feed together and the bot's grasp of its surroundings outstrips anything a lone sensor could deliver.
Obstacles, fixed or on foot, get spotted and identified fast thanks to the synchronised camera bank [[39]]. That built-in backup matters: when satellite fixes vanish inside an urban canyon, camera-based visual odometry and LiDAR SLAM take over to hold the machine's position.
02LiDAR: Drawing the Surroundings in Three Dimensions
For pavement work, LiDAR (Light Detection and Ranging) has a strong claim to being the single most important device. Laser pulses go out, the system clocks how long each one takes to return, and out of those timings comes a sharply detailed 3D point cloud.
A delivery bot puts LiDAR to work in several ways at once:
- Finding the footway: the boundary dividing pavement from roadway can be located with precision, so the machine keeps to its lane even where the edge is worn away or hidden.
- Fixing obstacle range: where cameras can only infer how far something sits, LiDAR hands over an exact figure, leaving no doubt about the gap to a passer-by or a hazard.
- Charting in 3D: a live three-dimensional model takes shape around the machine, tagging not merely what is present but precisely where it sits in space.
Anti-sunlight LiDAR now ships on current bots and holds its own even under fierce daylight [[35]]. That matters a great deal, since pavement work is mostly a daytime activity and solar glare would otherwise swamp the readings.
SLAM: Localisation and Mapping at the Same Time
Those LiDAR readings feed SLAM algorithms, through which the machine charts unfamiliar ground and keeps fixing its own place on the chart simultaneously. In dense city districts this is indispensable, because signals bouncing off tall buildings — the "urban canyon" — leave GPS unreliable on its own.
A familiar junction coming into view can prompt the bot to retrieve tactics that worked there before; a situation it has never met, by contrast, may trigger extra caution and, if needed, a human watching through teleoperation systems.
03Computer Vision: Giving Bots the Ability to "See"
LiDAR answers the question of where; computer vision takes on the question of what. Neural networks of considerable sophistication, trained on very large image sets, let delivery bots categorise the things around them on the fly.
The vision stack has to spot and sort the following:
| Object Category | Navigation Response | Priority |
|---|---|---|
| Pedestrians | Give way, hold a safe gap, forecast the path ahead | Critical |
| Crosswalks | Halt, let people pass, check the way is clear | Critical |
| Traffic Signals | Follow the walk and don't-walk signs at junctions | High |
| Static Obstacles | Route around them (fire hydrants, benches, trash cans) | Medium |
| Dynamic Obstacles | Forecast motion and reshape the course (e-scooters, bicycles, pets) | High |
| Sidewalk Edge | Keep a safe margin back from the kerb and roadway line | High |
Deep Learning Behind Object Recognition
Convolutional neural networks (CNNs) and transformer-style vision models now give delivery bots object-recognition performance on a par with human sight. The models have absorbed millions of images of city settings, learning to flag anything from a child's bike to a barrier thrown up for roadworks.
The camera stack draws on AI training methods much like those behind robotic grasping. As a result the system does more than locate an object: it reads its qualities — stiff versus yielding, parked versus moving, dangerous versus harmless.
Semantic Segmentation
Object detection alone is not the end of it. Semantic segmentation sorts every single pixel in the camera feed, which lets the bot tell passable ground (pavement slabs) off limits (lawns, roads, pools of water), even with no physical edge separating the two.
04Path Planning: The AI in Command
Sensors give the machine its picture, vision tells it what everything is, and next comes the question of how to get from here to there. That is the job of path-planning algorithms.
Path planning is the computation by which a machine works out a collision-free line from its starting point to a chosen destination [[30]]. For pavement bots, sheer distance is beside the point: the route has to be the safest, most efficient and socially acceptable one available.
Planning in Layers
Today's delivery bots plan in a hierarchy:
- Global planning: with GPS and pre-loaded maps, the broad course from restaurant or shop to the drop-off address takes shape, accounting for street layouts, the known footway network and the points where crossing is lawful.
- Local planning: the few metres immediately ahead (the next 5-10 meters) are replanned again and again as the bot moves, to dodge hazards, people and surprises — dozens of recalculations each second.
- Trajectory optimisation: rather than a mere geometric line, the machine shapes a trajectory that respects its own motion limits — acceleration, turning circle, speed caps — so movement stays smooth and steady.
Bots that are wider, slower and more cautious in their behaviour have the greatest need for robust routing approaches, according to research on dependable path planning for pavement delivery [[4]]. It helps explain the familiar sight of bots rolling at foot-traffic pace (3-5 mph) and sweeping well wide of hazards [[12]].
Getting Better with Every Trip
Machine learning lets the more advanced planners sharpen over time. Delivery by delivery, Coco Robotics and Avride are discovering what an unruly city demands of a navigator [[3]]. A stretch of footway jammed every lunchtime, a corner where satellite fixes always weaken — memories like these feed straight back into behaviour.
That is precisely where foundation model robotics enters: navigation models trained in advance can carry across cities and settings, cutting down the local map-building that used to precede deployment.
05Live Dodging of Moving Hazards
An empty pavement proves little. The real examination is the messy, shifting reality of foot traffic in a city, and meeting it calls for serious skill at dynamic obstacle avoidance.
On meeting a hazard, a delivery bot runs through the following:
- Pick it up through the sensor array
- Work out what kind of thing it is — a person, a vehicle, a fixed object
- Forecast where it is likely to be a moment from now
- Shape a substitute route that keeps clear of it
- Carry out the move steadily and without risk
Predictive Modelling
For things in motion, people above all, predictive models forecast future positions. Someone heading straight for the bot sets off a trajectory forecast and a matching course change — easing off speed, maybe, or drifting a touch to the right so the two can pass.
A proximity-seeking mechanism holds a safe gap to nearby hazards, while trace memory produces the best ways around them [[32]]. In practice the bot is never merely flinching at whatever sits under its nose: it keeps a working model of the whole scene and thinks several moves ahead.
Social Navigation
Not crashing is the floor, not the ceiling — pavement bots also have to move socially. In practice that means:
- Letting people go first, even where the bot technically holds right of way
- Holding sensible gaps when following others
- Making intentions plain, through movement or indicator lights
- Steering clear of abrupt, jerky motion that could give anyone a fright
Those social rules are baked straight into the planner, so a bot neither collides with the public nor turns into a nuisance or a hazard.
Coping with Edge Cases
Even advanced AI meets moments it cannot resolve alone. A genuinely ambiguous scene — a footway sealed off end to end by roadworks, say — gets passed by the bot to a remote operator through teleoperation systems, who supplies the human judgement it lacks [[1]].
06The Hard Parts of Getting Around a City
Urban footways throw up distinctive difficulties that press against the edges of what autonomous navigation can do. They also account for the slow, deliberate pace at which delivery bots have been rolled out.
| Challenge | Impact | Solution |
|---|---|---|
| GPS Signal Loss | Satellite signals get cut off in urban canyons | LiDAR SLAM together with visual odometry for dead reckoning |
| Poor Lighting | Nightfall and shadows blunt the cameras | LiDAR plus radar — sensors that do not depend on light |
| Adverse Weather | Rain, snow and fog wear down sensor performance | Redundant sensors alongside algorithms tuned for weather |
| Crowded Sidewalks | Dense foot traffic makes navigation genuinely complicated | Predictive models paired with conservative speed caps |
| Construction Zones | Obstacles that were not there yesterday, routes redrawn | Live mapping with teleoperation held in reserve |
| Accessibility | As discussions of robotics challenges point out, getting across curb cuts, ramps and uneven ground | Advanced suspension plus classification of the terrain |
The Security Side of the Problem
Physical navigation is only one front; cyber threats have to be fended off as well. In much the same way that grasping AI deepfake detection is what shields people from tampered media, an autonomous bot has to confirm its feeds have not been spoofed or compromised, since corrupted perception could push it into dangerous mistakes.
07Where Sidewalk Navigation Is Headed Next
The technology carrying delivery bots along city footways keeps moving fast. A handful of emerging developments look set to make the systems both more capable and more common:
Navigation Powered by Foundation Models
Vast and steadily growing libraries of video are being put to work pre-training navigation foundation models for autonomous pavement travel [[27]]. Fed on millions of hours of recorded urban navigation, such models could let a bot carry its skills into a fresh city with very little extra training.
V2X, or Vehicle-to-Everything, Communication
Bots of the future may talk directly to the traffic fabric — smart lights, crossing signals — and to other autonomous vehicles. Out of those exchanges would come a cooperative navigation system, safer and more efficient than each machine deciding in isolation.
Sharper Simulation and Digital Twins
A new city no longer has to be learned entirely on its streets. Before launch, bots can now log heavy training hours inside photorealistic simulations, digital twins of the actual place, and pick up its local oddities — that strange junction downtown, the farmer's market that is always packed — without anyone being put at risk.
As the countries leading AI robotics demonstrate, heavy national investment in smart-city fabric is producing surroundings in which delivery bots can flourish: lanes set aside for them, better-kept footways and traffic management joined up end to end.
Swarm Intelligence
Several bots at work in one district can exchange navigation data on the spot. The moment one finds a sealed footway or a fresh hazard, every bot nearby can know at once, and that shared perception lifts both the efficiency and the safety of the fleet as a whole.