011产品思考

如果汽车可以自动驾驶，为什么自行车不行？

从一次骑行开始，关于 AI、世界模型和未来交通的一些不成熟思考。

2026/10/25约 15 分钟

智能交通
世界模型
骑行
产品思考

01｜骑车的时候，我突然想到一个问题

有时候，我觉得自己提出的问题还挺奇怪的。

比如，骑自行车的时候，我突然想到：

既然汽车可以自动驾驶，那为什么自行车不行？

这个问题听起来似乎有点多余。

毕竟，自行车本来就是用来骑的。如果连骑车这件事都交给 AI，好像多少失去了一点意义。

可仔细想想，汽车不也是用来开的吗？

我们为什么愿意花费大量时间和资源，让汽车学会感知环境、识别障碍物、预测其他车辆的行为，甚至在特定条件下替代人类完成驾驶，却很少讨论自行车是否需要类似的能力？

也许是因为，自行车太简单了。

两个轮子，一个车把，一套刹车系统。它似乎不需要那么复杂的智能。

但问题恰恰出现在这里。

如果我们研究智能驾驶的最终目的是提高出行的安全性和效率，那么这种能力是否应该只属于汽车？

我开始重新观察自己骑车时遇到的情况。

比如，路边突然打开的车门，转弯时没有注意到的行人，路口突然驶出的电动车，以及被车辆挡住的视线。

很多时候，危险并不是因为我不知道自己要去哪里，而是因为我无法提前知道下一秒会发生什么。

这和自动驾驶汽车需要解决的问题，似乎有一些相似之处。

02｜打开高德地图，算不算 AI 辅助驾驶？

骑车的时候，我有时会打开导航。

它可以告诉我应该在哪个路口转弯，前方的路线怎么走，以及预计还有多久到达目的地。

那这算不算一种 AI 辅助驾驶？

后来我发现，需要先区分三个概念。

导航、驾驶辅助和自动驾驶。

导航主要解决的是「我应该走哪条路」。

驾驶辅助试图帮助使用者更安全地完成某些操作，例如识别潜在危险、提供预警。

自动驾驶则更进一步，需要系统在特定条件下持续承担感知、决策与车辆控制任务。

这几件事情可以相互关联，却并不相同。

高德地图即使使用了智能算法进行路径规划，也不意味着它正在替我控制自行车。

它知道我应该在哪里转弯，却未必知道此刻有没有一个小孩突然从旁边跑出来。

如果把一部带有摄像头、定位系统和运动传感器的智能手机固定在自行车上，理论上，我们又能获得哪些新的能力？

它或许可以尝试识别周围的车辆和行人、记录骑行轨迹，或者通过声音提供一定的环境提醒。

但能够识别危险，不代表能够可靠地避免危险。

普通手机的摄像头有视野限制，定位可能存在误差，复杂的光照和天气也会影响感知结果。

更重要的是，手机没有经过验证的车辆控制系统，无法因为判断前方有危险，就可靠地完成刹车、转向和保持平衡。

把手机装在车上，不等于给自行车安装了自动驾驶系统。

而未经验证的实验功能，也不应该在真实道路上被当作安全保障。

这让我开始意识到：

从「知道周围发生了什么」，到「能够安全地采取行动」，中间可能隔着比想象中更远的距离。

03｜自行车真的不能自动驾驶吗？

答案其实不是不能。

2026 年 10 月，我看到一个很有意思的研究案例。

苏黎世联邦理工学院（ETH Zurich）的学生团队开发了一款名为 AutoBike 的自主行驶自行车原型。

它并不是为了让骑车的人在路上放开双手。

相反，它试图解决一个非常具体的问题：共享自行车在不同地点之间的调度。

比如，某个区域停满了共享自行车，而另一个区域却没有足够的车可用。

传统方式需要工作人员把车辆重新运输到需要的地方。

AutoBike 则尝试让车辆在没有骑乘者的情况下，自主前往目标位置。

为了实现这个目标，团队为它设计了环境感知、定位、路径规划、转向和制动系统。

有意思的是，它还配备了可以收起的辅助轮。

当人正常骑车时，辅助轮收起；当自行车自主移动时，辅助轮放下，以提高低速行驶时的稳定性。

它并没有执着于让一辆普通自行车像人一样保持平衡，而是根据实际任务，重新设计了技术方案。

这件事让我觉得很有启发。

我们总以为技术创新是在不断增加能力。

让汽车更聪明，让机器人更灵活，让 AI 能够完成更多任务。

但有时，真正重要的可能是重新理解任务本身。

为什么一定要让自行车学会像人一样骑行？如果我们真正想解决的是车辆调度问题，那么让它安全地自主移动就够了。

这和做 AI 产品有一点相似。

不是先决定我要使用多么先进的技术，再寻找能够展示它的场景。

而是先想清楚，我究竟希望解决什么问题。

04｜比自行车自动驾驶更重要的，可能是自行车的安全

不过，AutoBike 解决的是车辆调度，并不是普通人在道路上骑行时的安全问题。

如果回到最开始的疑问：我真的需要一辆自动驾驶自行车吗？

我发现自己的答案可能是，不一定。

我并不希望自行车替我完成所有操作。

骑车本身就有乐趣。

我喜欢自己控制方向，感受速度，选择路线，以及偶尔漫无目的地探索一座城市。

相比让自行车自己骑，我可能更希望它能提前告诉我：

前面有什么我还没注意到的危险？

旁边的汽车是不是准备右转？

前方被遮挡的路口，是否存在突然出现车辆的可能？

当我的注意力没有覆盖到某个方向时，系统能不能提供适当的提醒？

从这个角度来看，我真正需要的或许不是「自动驾驶」，而是某种更可靠的「智能骑行辅助」。

甚至不一定需要接管自行车。

它可以只是一个足够克制的提醒系统。

在必要的时候告诉我危险，在不必要的时候保持安静。

当然，这里又出现了新的问题。

如果系统误报太多，我可能会逐渐忽略它。

如果提醒太晚，它可能根本无法提供帮助。

如果它给出了错误的安全暗示，反而可能让我降低警惕。

所以，智能骑行辅助不是简单地把摄像头识别结果转换成语音提醒。

它涉及人的反应时间、注意力、环境感知、风险判断，以及人与系统之间如何分配责任。

有时，一个不可靠的智能功能，甚至可能不如没有这个功能。

这让我开始觉得，技术的价值不一定来自它能够替代人类多少操作。

有时候，让人保有控制权，同时获得更好的信息，本身就是一种进步。

05｜如果真正需要变聪明的不是自行车，而是道路呢？

思考到这里，我突然发现，自己可能一直在用一种很自然的方式理解交通智能化。

汽车不够安全，就给汽车增加传感器。

自行车无法感知危险，就给自行车安装摄像头。

行人过马路存在风险，就设计更智能的提醒系统。

我们似乎习惯了以单个交通参与者为单位，不断提高每个个体的智能程度。

但交通系统并不是一个个独立运行的设备。

它更像是一个不断变化的复杂网络。

一辆汽车的行为会影响旁边的自行车。

一名行人的突然转向可能改变多辆车的运动轨迹。

一处被遮挡的路口，则可能让所有参与者都失去一部分信息。

那么问题来了：

为什么所有车辆都要独自承担感知整个交通环境的任务？

如果道路本身也能够提供信息呢？

例如，在视线被建筑物遮挡的路口，路侧设备发现了一个正在接近的骑行者。

这条信息能否提前传递给准备右转的汽车？

如果车辆和道路之间可以交换可信的交通状态信息，它们是否有机会发现单个摄像头无法看到的危险？

这种方向在智能交通研究中被称为车路协同，相关技术包括 V2X，也就是车辆与其他车辆、道路基础设施等进行信息交互。

2026 年，卡内基梅隆大学的一项研究就讨论了如何利用协同感知与轨迹预测，更好地保护行人和骑行者等弱势道路使用者。

它让我想到一个有意思的区别。

单车智能尝试回答：

「我能看到什么，我应该怎么做？」

道路协同则进一步尝试回答：

「我们各自看到了什么，能不能共享有用的信息，从而一起减少风险？」

当然，协同系统也有自己的问题。

比如通信延迟、网络覆盖、设备成本、隐私保护，以及不同参与者的信息是否可信。

因此，它不可能简单替代车辆自身的感知和判断。

但我越来越觉得，两者并不是非此即彼的关系。

单个设备需要有独立应对风险的能力，而道路系统也可以尝试提供更多信息，弥补个体感知的局限。

真正值得追求的，也许不是让每辆车都无所不知，而是让整个交通系统不必只依赖某一辆车的判断。

06｜世界模型能不能提前知道危险？

顺着这个问题，我又想到了最近经常被讨论的世界模型。

如果说传统的环境感知主要关注「现在发生了什么」，那么世界模型试图进一步表示环境如何变化，以及不同动作可能带来怎样的后果。

当然，不同研究对世界模型的定义并不完全相同。

但我最感兴趣的是其中关于预测的部分。

假设一辆汽车正准备右转，旁边有一个骑自行车的人。

系统不仅需要知道那里有一辆自行车，还需要考虑：

骑行者会不会继续直行？

汽车转弯的轨迹是否会与自行车发生冲突？

如果骑行者突然避让一个障碍物，接下来可能出现什么情况？

这些问题并不只是目标识别，而涉及对未来状态的预测。

不过，未来不是一条确定的轨迹。

一个骑行者可能直行，也可能减速，还可能突然转向。

再先进的模型，也无法保证准确预测每一个人的行为。

所以我开始怀疑，我们是否真的应该追求某种「足够准确就能避免所有事故」的预测系统。

也许更合理的目标，是让系统能够识别多种可能的未来，并且知道自己什么时候没有把握。

当某种后果虽然发生概率不高，却可能造成严重伤害时，系统是否应该保留更大的安全余量？

当预测不确定性过高时，是否应该优先减速，而不是继续相信最有可能发生的那一种情况？

从这个角度看，智能交通中的世界模型不应该只是一个会预测未来的模型。

它还需要与决策、控制、安全约束和实际验证结合。

预测只是其中的一环。

而且预测结果越复杂，越需要思考如何验证它是否可靠。

真正困难的，可能不是让 AI 想象未来，而是让它在无法确定未来时，依然采取足够安全的行动。

07｜一辆电动轮椅，又应该属于哪一种交通工具？

在这个过程中，我还想到了一个更容易被忽略的对象。

电动轮椅。

从外观来看，它有电机、有轮子，也能够自主移动。

但它的使用目的、运行环境和安全需求，又与汽车、自行车存在很大不同。

我不想简单地根据它有没有电机，就把它归入某一种交通工具。

从技术和产品设计的角度，我更关心的是它需要什么样的能力。

假设一位行动不便的人，需要使用电动轮椅经过一段不平整的人行道。

他可能真正需要的，并不是让轮椅完全自动驾驶。

而是辅助避障、缓慢通行、识别台阶或坡度，以及在危险情况下可靠地停止。

如果把汽车自动驾驶的整套技术直接搬过来，成本和复杂度可能远远超过实际需要。

但如果因为轮椅速度不快，就忽略它的智能辅助需求，也显然不合理。

所以我开始觉得，判断一种交通工具需不需要 AI，或许不应该首先看它属于机动车还是非机动车。

而应该看：

它的使用者是谁？

实际面临什么风险？

哪些操作适合交给机器？

哪些决策必须由人保留？

以及技术介入之后，是否真的能改善安全性和自主行动能力？

从汽车到自行车，再到电动轮椅，我们可能需要的并不是同一种自动驾驶。

而是针对不同出行方式的智能辅助能力。

08｜未来的交通，应该让谁变得更聪明？

最开始，我只是觉得骑自行车的时候也拥有一点智能驾驶能力，似乎很有意思。

后来发现，这个问题背后涉及感知、控制、机器人、世界模型、人机交互以及城市交通系统。

而且，越往下想，越觉得没有那么容易回答。

我不是自动驾驶领域的研究者，也没有足够的专业积累来判断哪一种技术路线一定更好。

目前的这些想法，更多来自日常观察、查阅资料和对不同应用场景的比较。

但我很喜欢这个思考过程。

因为它让我意识到，一个看起来很小的问题，有时会通往一个完全不同的研究方向。

我最初问的是：

「为什么自行车不能自动驾驶？」

后来变成：

「自行车真的需要自动驾驶吗？」

再后来，我开始关心：

「未来的交通智能化，究竟应该优先提高单个交通工具的智能程度，还是整个道路系统的安全协同能力？」

现在，我更倾向于认为，这可能不是一道二选一的题。

车辆需要具备独立应对危险的能力，道路系统也可以通过协同提供更多的信息。

技术应该根据具体场景选择合适的介入程度，而不是让所有交通工具朝着完全自动驾驶的方向发展。

或许有一天，自行车真的可以在城市里自己行驶。

也可能那时，我们依然会选择亲自骑车。

只是道路能够更早发现危险，汽车能够更好地理解骑行者的意图，智能设备能够在必要时给出帮助。

而那些需要额外支持的人，也能够更加安全、独立地抵达自己想去的地方。

如果最终能够做到这些，那么自行车究竟有没有实现完全自动驾驶，好像也没有那么重要了。

我想要的不是一辆什么都替我做的自行车。

而是一个即使我只是普通地骑着车，也能让我更安心的世界。

2026 年 10 月 · 一次关于智能交通的好奇

本文是一篇个人观察与探索笔记，涉及的技术设想不代表已经完成工程验证。

01 | A question that occurred to me while cycling

Sometimes I think the questions I ask are a little strange.

For example, while riding my bike, I suddenly wondered:

If cars can drive themselves, why can’t bicycles?

The question might sound unnecessary.

After all, a bicycle is meant to be ridden. Handing even that over to AI seems to take something away from it.

But aren’t cars meant to be driven, too?

Why are we willing to spend so much time and so many resources teaching cars to perceive their surroundings, recognize obstacles, predict other vehicles’ behavior, and even drive instead of a person under certain conditions, yet rarely discuss whether bicycles need similar abilities?

Perhaps bicycles seem too simple.

Two wheels, a handlebar, a braking system. They don’t appear to need much intelligence.

But that is precisely where the question begins.

If the ultimate purpose of intelligent driving is safer and more efficient travel, should those capabilities belong only to cars?

I began paying closer attention to what I encounter while cycling.

A car door opens unexpectedly. A pedestrian goes unnoticed during a turn. An electric scooter emerges from an intersection. A parked vehicle blocks my view.

Often, the danger isn’t that I don’t know where I’m going. It is that I can’t know what will happen in the next second.

That seems to have something in common with the problems autonomous cars need to solve.

02 | Does opening Amap count as AI-assisted driving?

Sometimes I use navigation while cycling.

It tells me which intersection to turn at, how the route continues, and roughly how long it will take to arrive.

Does that count as AI-assisted driving?

I eventually realized that I needed to separate three concepts first.

Navigation, driving assistance, and autonomous driving.

Navigation mainly answers, “Which route should I take?”

Driving assistance tries to help someone perform certain tasks more safely, such as identifying a potential hazard and issuing a warning.

Autonomous driving goes further: under specified conditions, the system must continuously handle perception, decisions, and vehicle control.

These functions can be related, but they are different.

Even if Amap uses intelligent algorithms to plan a route, that doesn’t mean it is controlling my bicycle.

It knows where I should turn. It may not know whether a child is about to run out beside me.

If I mounted a smartphone with a camera, positioning, and motion sensors on a bicycle, what new capabilities might become possible?

It might try to recognize nearby vehicles and pedestrians, record my route, or provide some spoken information about the surroundings.

But recognizing a hazard doesn’t mean reliably avoiding it.

An ordinary phone camera has a limited field of view. Positioning can be inaccurate, and difficult lighting or weather can affect perception.

More importantly, the phone has no validated vehicle-control system. Recognizing danger ahead doesn’t let it reliably brake, steer, and keep the bicycle balanced.

Mounting a phone on a bicycle does not turn it into an autonomous vehicle.

Unvalidated experiments should not be treated as safety protection on real roads, either.

This made me realize:

The distance between knowing what is happening and being able to act safely may be much greater than it first appears.

03 | Is an autonomous bicycle actually impossible?

It isn’t.

In October 2026, I came across an interesting research project.

A student team at ETH Zurich developed an autonomous bicycle prototype called AutoBike.

It wasn’t intended to let riders take their hands off the handlebars on the road.

Instead, it addressed a specific task: redistributing shared bicycles between locations.

One area might have too many bikes while another has too few. Usually, workers have to transport them to where they are needed. AutoBike explores letting an unoccupied bicycle travel to a destination itself.

The team developed perception, localization, planning, steering, and braking systems. It also fitted retractable support wheels: raised for ordinary riding, lowered for stability during autonomous movement at low speed.

Rather than insisting that an ordinary bicycle learn to balance like a person, the team adapted the technical approach to the actual task.

I found that encouraging.

We often imagine innovation as an endless expansion of capabilities: smarter cars, more agile robots, AI that can do more things.

Sometimes, though, the more important step is understanding the task differently.

Why must a bicycle learn to ride like a person? If the problem is redistribution, safe autonomous movement may be enough.

There is a parallel with building AI products.

I shouldn’t choose an impressive technology first and then look for somewhere to demonstrate it.

I need to understand what I am trying to solve first.

04 | Bicycle safety may matter more than bicycle autonomy

AutoBike addresses redistribution, though, rather than the safety of someone riding on a road.

Returning to my original question, do I actually need a self-driving bicycle?

My answer might be: not necessarily.

I don’t want a bicycle to do everything for me.

Riding is enjoyable in itself.

I like choosing my direction, feeling the speed, finding a route, and sometimes exploring a city without a particular destination.

Instead of a bicycle that rides itself, I might prefer one that can tell me:

What danger ahead have I not noticed yet?

Is the car beside me about to turn right?

Could a vehicle suddenly emerge from that obscured junction?

When my attention doesn’t cover a particular direction, could the system offer a useful warning?

What I want may be a more reliable form of intelligent cycling assistance.

It might not need to take control of the bicycle at all.

A restrained warning system could be enough.

Warn me when necessary, and stay quiet when it isn’t.

Of course, that raises further questions.

If it produces too many false alarms, I may gradually ignore it.

If it warns me too late, it may offer no help.

If it wrongly implies that everything is safe, it may make me less attentive.

Intelligent cycling assistance therefore requires more than converting camera detections into spoken warnings.

It involves reaction time, attention, perception, risk judgment, and the division of responsibility between a person and a system.

Sometimes an unreliable intelligent feature could be worse than having no such feature.

This makes me think that technology’s value needn’t be measured only by how many human actions it replaces.

Keeping people in control while giving them better information can be progress in itself.

05 | What if the road needs to become smarter, too?

At this point, I noticed a habit in the way I was thinking about intelligent transport.

A car isn’t safe enough, so we add sensors to the car.

A bicycle can’t detect hazards, so we add a camera to the bicycle.

Crossing the road is risky, so we design a smarter warning system for pedestrians.

We tend to treat each road user as a separate unit and keep improving each one’s intelligence.

Yet a transport system isn’t a collection of independently operating devices.

It is more like a changing, interconnected network.

A car’s behavior affects a bicycle beside it.

A pedestrian’s sudden change of direction may alter the paths of several vehicles.

An obscured junction can leave everyone with incomplete information.

That raises another question:

Why should every vehicle have to perceive the entire traffic environment on its own?

What if the road could provide information as well?

For example, roadside equipment at a junction obscured by a building might detect an approaching cyclist.

Could that information reach a car preparing to turn right before the driver sees the cyclist?

If vehicles and infrastructure exchanged trustworthy traffic-state information, could they notice hazards that a single camera cannot see?

This direction is often described as vehicle–infrastructure cooperation. Related technologies include V2X: communication between vehicles and other vehicles, infrastructure, and other participants.

A 2026 Carnegie Mellon University report discusses cooperative perception and trajectory prediction for the safety of vulnerable road users, including pedestrians and cyclists.

It suggested an interesting distinction to me.

Intelligence within a single vehicle asks:

“What can I see, and what should I do?”

Cooperation asks a further question:

“What have we each seen, and can sharing useful information help us reduce risk together?”

Cooperative systems have problems of their own, of course.

Communication latency, network coverage, equipment costs, privacy, and whether other participants’ information is trustworthy all matter.

They cannot simply replace a vehicle’s own perception and judgment.

But I increasingly think these approaches can work together.

Individual devices need the ability to respond to risk independently. The road system can also try to provide additional information that addresses the limits of individual perception.

Perhaps the goal is less about making every vehicle all-knowing, and more about making the transport system less dependent on any single vehicle’s judgment.

06 | Could a world model anticipate danger?

Following this question led me to world models, which have been discussed a lot recently.

If conventional perception focuses mainly on what is happening now, a world model tries to represent how an environment changes and what consequences different actions might have.

Definitions vary across research, of course.

The part that interests me most is prediction.

Imagine a car preparing to turn right with a cyclist beside it.

The system needs to do more than recognize that a bicycle is there. It must consider:

Will the cyclist continue straight ahead?

Will the car’s turning path conflict with the bicycle’s path?

What might happen if the cyclist suddenly swerves around an obstacle?

These questions concern possible future states, beyond object recognition alone.

But the future isn’t a single certain trajectory.

A cyclist might go straight, slow down, or turn suddenly.

Even an advanced model cannot guarantee an accurate prediction of everyone’s behavior.

I therefore wonder whether pursuing a prediction system that could “avoid every accident once it is accurate enough” is the right objective.

Perhaps a more reasonable goal is to represent several possible futures and recognize when confidence is low.

If an outcome is unlikely but could cause serious injury, should the system leave a larger safety margin?

When uncertainty is high, should it prioritize slowing down instead of continuing to trust the most likely scenario?

From this perspective, a world model for intelligent transport needs more than an ability to predict.

It needs to work with decisions, control, safety constraints, and practical validation.

Prediction is one part of that process.

And the more complex the predictions become, the more carefully we need to consider how to validate their reliability.

The harder problem may be getting AI to act safely when the future is uncertain, rather than getting it to imagine the future.

07 | Where does a powered wheelchair fit?

Along the way, I thought of another vehicle that is easier to overlook.

A powered wheelchair.

It has motors and wheels and can move under its own power.

Yet its purpose, operating environment, and safety requirements differ considerably from those of cars and bicycles.

I don’t want to classify it simply by whether it has a motor.

From a technical and product-design perspective, I care more about the capabilities its user needs.

Imagine someone with limited mobility using a powered wheelchair along an uneven pavement.

They may not need fully autonomous travel.

They might need help avoiding obstacles, moving slowly, recognizing steps or slopes, and stopping reliably when something is dangerous.

Transplanting an entire autonomous-car system could introduce far more cost and complexity than the situation requires.

Yet dismissing intelligent assistance because a wheelchair moves slowly would be unreasonable, too.

Perhaps the first question in deciding whether a vehicle needs AI should be about its users and tasks, rather than whether it falls into a motorized or non-motorized category.

Who is using it?

What risks do they actually face?

Which actions are suitable for a machine to perform?

Which decisions must stay with the person?

And does introducing the technology actually improve safety and independent mobility?

Cars, bicycles, and powered wheelchairs may need different kinds of intelligent assistance, rather than the same form of autonomous driving.

08 | Who should become smarter in the future of transport?

I began by thinking it would be interesting to have a little intelligent-driving capability while riding a bike.

I eventually found myself considering perception, control, robotics, world models, human–machine interaction, and urban transport systems.

The more I thought about it, the harder the question became to answer.

I am not an autonomous-driving researcher. I don’t have the professional experience to decide which technical approach must be better.

These thoughts come mainly from everyday observations, reading, and comparing different situations.

But I enjoy the process.

A small question can sometimes open an entirely different direction of inquiry.

My first question was:

“Why can’t bicycles drive themselves?”

It became:

“Do bicycles actually need to drive themselves?”

Later, I began asking:

“Should intelligent transport prioritize smarter individual vehicles, or safer cooperation across the road system?”

I now suspect that we don’t have to choose only one.

Vehicles need the ability to respond to danger independently. Cooperation within the road system can provide more information as well.

Technology should intervene to a degree appropriate to the situation, rather than push every kind of vehicle toward full autonomy.

Perhaps bicycles will one day travel through cities on their own.

Perhaps we will still choose to ride them ourselves.

The difference might be that roads can detect hazards earlier, cars can better understand cyclists’ intentions, and intelligent devices can help when needed.

People who need additional support might also reach their destinations more safely and independently.

If we could achieve that, whether bicycles have reached full autonomy might no longer seem so important.

I don’t want a bicycle that does everything for me.

I want a world where I can feel safer while simply riding one.

October 2026 · A curiosity about intelligent transport

This is a personal observation and exploration note. The technical ideas proposed in it should not be taken as having completed engineering validation.

不是每件事都需要一个标签。

探索线索