ADM-201 dump PMP dumps pdf SSCP exam materials CBAP exam sample questions

为什么自动驾驶汽车要花这么长时间? – 译学馆
未登录,请登录后再发表信息
最新评论 (0)
播放视频

为什么自动驾驶汽车要花这么长时间?

Why Are Self-Driving Cars Taking So Long?

科学秀由Brilliant.org赞助
SciShow is supported by Brilliant.org.
目前 你可能听说过自动驾驶马上就要来了
By now, you’ve probably heard that self-drivingcars are coming soon.
如果你还没听说过——大惊喜!
If you haven’t—surprise!
真的很快就要来了!
They’re coming soon!
但是人们已经有近10年都在这么说
But people have been saying that for at leasta decade,
我仍然不相信
and I still can ’ t buy a
我在副驾驶打盹儿
car that ’ ll drive me to work
汽车会自动把我带到工作地点
while I nap in the passenger seat.
有的车已经有部分自动驾驶系统
Some cars already come with partial autonomy,
比如特斯拉的自动驾驶
systems like Tesla ’ s Autopilot,
辅助司机驾驶有时甚至能够代理驾驶
that assist drivers or sometimes even take control.
但是如果有危险
But they still need a human
还需要人类司机
driver who can grab the reins on short notice
快速接管
if things get dicey,
这也是今年早些时候
which is why someone in the UK got arrested
英国一名司机坐到副驾驶上而被捕的原因
earlier this year for trying the passengerseat thing.
未来几年很可能会
There are some fully driverless vehicles thatmight be released
有完全无人驾驶的车辆上市
in the next few years,
但是只适用于某些特殊用途
but they ’ re only meant for very specific uses,
比如长途运输 限于某些街道或街区的出租车
like long-haul trucking or taxis confinedto certain streets and neighborhoods.
这是因为广泛适用的自动驾驶很难
That’s because general-purpose driving ishard!
所需软件必须设计得能处理
The software has to work out a lot
许多棘手的问题
of really tricky questions
把传感器上的信息转换成命令控制方向盘和踏板
to turn information from its sensors into commands to the steering and pedals.
尽管对这项研究
And despite all the money and brainpower
投入了许多人力 物力
that ’ s being poured into research,
仍然存在很多大问题
there are still major challenges at everystep along that path.
第一个就是自动驾驶汽车
The first thing a self-driving car has to
必须判断出周围的事物
do is figure out what ’ s around it,
以及这些事物所在的位置
and where everything is.
被称为感知阶段
It’s called the perception stage.
人类看一眼就能做到这一点
Humans can do this at a glance, but a carneeds
而汽车需要汇聚多种传感器的信息:相机 雷达超生波感应器
a whole cornucopia of sensor data: cameras,radar, ultrasonic sensors,
和激光定位器——使用激光而不是无线电
and lidar, which is basically detailed 3D
构造三维图像
radar that uses lasers instead of radio.
目前汽车确实需要解释所有这些数据
Today’s autonomous vehicles do pretty wellat interpreting all that data
来得到周围道路 汽车 交通信号灯等情况的
to get a 3D digital model of their surroundings
三围数字模型
the lanes, cars, traffic lights, and so on.
但是软件辨别物体通常是很难的
But it’s not always easy to figure out what’swhat.
例如 如果许多物体靠得很近
For example, if lots of objects are closetogether say,
比如 在一大堆人群中
in a big crowd
软件很难将人们分开
of people it ’ s hard for the software to separate them.
所以在大城市这样的行人密集区域
So to work properly in pedestrian-packed areaslike major cities,
汽车要正常运转 考虑的不仅仅是当前的影像
the car might have to consider not just thecurrent image
还要考虑过去的几秒钟的内容
but the past few milliseconds of context,too.
这样 软件才能够
That way, it can group
将一团一起运动的点
a smaller blob of points moving together
辨别为在街上走的行人
into a distinct pedestrian about to step intothe street. Also,
有的东西的确对计算机来说很难辨别
some things are just inherently hardfor computers to identify:
漂浮的塑料袋对传感器来说
a drifting plastic bag looks just
就是一个实心的 较重的
as solid to the sensors as a heavier,
危险的 装满垃圾的袋子
and more dangerous, bag full of trash.
这种混淆会导致不必要的刹车
That particular mix-up would just lead tounnecessary braking,
而错误的判断又会是致命的
but mistaken identities can be fatal:
在2016年特斯拉致命撞车事件中
in a deadly Tesla crash in 2016,
自动驾驶相机误把卡车的侧面当成了明亮的天空
the Autopilot cameras mistook the side of a truck for washed-out sky.
你还需要确保系统是可靠的
You also need to make sure the system is dependable,
即使有意外情况发生
even if there are surprises.
例如 如果相机出错了
If a camera goes haywire, for example,
汽车要能够叠加各种信息
the car has to be able to fall back on overlapping sources of information.
还需要足够了解死臭鼬
It also needs enough experience to learn aboutdead skunks,
多人自行车 反铲挖掘机
conference bikes, backhoes sliding off trucks,
可能会在路上出现的奇怪的东西
and all the other weird situations that might show up on the road.
学者们经常会用运行游戏“侠盗猎车手”中的模拟器
Academics often resort to running simulationsin Grand Theft Auto yes,
对 侠盗猎车手
that Grand Theft Auto.
有些公司有更复杂的模拟器
Some companies have more sophisticated simulators,
但是即使想到这么多 还有设计者想不到的地方
but even those are limited by the designers’imaginations.
仍然有一些情况是很难感知到的
So there are still some cases where perceptionis tricky.
真正难以处理的问题是第二个阶段:预测
The really stubborn problems, though, comewith the next stage: prediction.
知道当前行人在哪里
It ’ s not enough to know where
其他司机在哪里 是不够的
the pedestrians and other drivers are right now
汽车要预测接下来往哪儿开
the car has to predict where they’re goingnext
接下来是第三部计划:自己的移动路线
before it can move on to stage 3: planning its own moves.
有时预测是简单的
Sometimes prediction is straightforward:
汽车的右灯闪烁表明它要向右并道
a car ’ s right blinker suggests it ’ s about to merge right.
这是预测很容易
That’s where planning is easy.
但是有时候计算机不懂它们的人类主人
But sometimes computers just don’t get theirhuman overlords.
比如说 一辆车靠近 减速 打闪灯
Say an oncoming car slows down and flashes its lights
当你在左车道等待时
as you wait for a left.
这时转弯是安全的
It ’ s probably safe to turn,
但是这件事情对计算机来说很难理解
but that ’ s a subtle thing for a computer to realize.
让预测变得复杂的是
What makes prediction really complicated,though,
转弯的安全性
is that the safety
不是你认出来的
of the turn isn ’ t something you just recognize
而是谈判的结果
it’s a negotiation.
如果你要左转 车向前挪
If you edge forward like you ’ re
对方司机会做出反应
about to make the left, the other driver will react.
所以这是预测和计划的反馈循环
So there’s this feedback loop between predictionand planning.
实际上 研究人员已经发现
In fact, researchers have found
当你要并入快车道时
that when you ’ re merging onto the highway,
如果你不看其他人的反应
if you don’t rely on other people to reactto you,
你可能永远不能安全的驶入
you might never be able to proceed safely.
所以自动驾驶汽车如果不能够判断
So if a self-driving car isn ’ t assertive enough,
它的进展就会受阻
it can get stuck:
所有的动作似乎都是不安全的
all actions seem too unsafe,
你就有了研究者称为“凝固机器人问题”
and you have yourself what researchers call the “ freezing robot problem. ”
它自身也是不安全的
Which itself can be unsafe!
程序员主要采用两种方法解决这种问题
There are two main ways programmers try to work around all this.
一种是让汽车独自
One option is to have the car think
根据其他车的行为
of everyone else ’ s actions
作出自己的动作
as dependent on its own.
但是这可能会导致过度激进行为 这也是很危险的
But that can lead to overly aggressive behavior, which is also dangerous.
同在路上开车的人
People who drive that way are
可能会在路上迂回
the ones who end up swerving
穿梭于各个车辆之间
all over the highway trying to weave betweenthe cars.
顺便说一句 千万别这样
Don’t do that, by the way.
另一种方法是让汽车预测每个人的各种行为
Another option is to have the car predicteveryone’s actions collectively,
就好像自己和其他车一样
treating itself as just one more car interacting
彼此互动
like all the rest,
然后做出最适合当时情形的举动
and then do whatever fits the situation best.
这种方法的问题是
The problem with that approach is
你必须极度简化事情 迅速做出决定
that you have to oversimplify things to decide quickly.
在众多问题:判断周围有什么物体
Finding a better solution to prediction andplanning
理解其他司机在做的事情
is one of the biggest unsolved problems inautonomous driving.
知道如何做出反应之中
So between identifying what’s around them,
找到较好的方法预测和计划
interpreting what other drivers will do,
是自动驾驶要解决的最大问题
and figuring out how to respond,
有许多类型的自动驾驶汽车
there are a lot of scenarios self-driving cars
都没有完全解决这个问题
aren ’ t totally prepared for yet.
这并不意味着无人驾驶车不会很快上路
That doesn’t mean driverless cars won’thit some roads soon.
还有更多简单的情形
There are plenty of more straightforward situations
你不会遇到这类复杂问题的
where you just don’t encounter these typesof problems.
但是对于在任何地方都可以开的自动驾驶汽车
But as for self-driving cars that can go anywhere…
我们只能说工程师
let ’ s just say the engineers won ’ t be
不会很快失业
out of a job any time soon.
我喜欢这种带有某种解决问题的思考水平
I love the layers of thinking involved in this kind of problem solving.
虽然我不是设计自动驾驶汽车的工程师
And while I’m not an engineer designingself-driving cars,
我还是要在Brilliant.org做这种思考
but I still get to practice this kind of thinking on Brilliant.org.
现在我在学习卷积神经网络课程
Right now, I’m working through the ConvolutionalNeural Networks lesson
来帮我学习如何使用神经网络
to help me learn how to work with neural networks.
我已经学完了概述部分
I’ve already gone through the overview,
这个“应用与实践”测试上有一个汽车图标
and this “Applications and Performance”quiz has a car on it,
所以下次我要试一下那个
so that’s what I’m going to try my handat next.
这个测试解释了网络是如何工作的
The quiz already explained how this networkworks.
然后问我们如何修正
And then it ’ s asking how we
让它符合这个Imagenet难题
should modify it to suit this imagenet challenge,
帮助它更好的对事物进行分类
to help it categorize objects better.
我认为答案是C
I think the answer is C:
在终端加上一个完全连接的网络
to add a fully connected network at the end
帮助预测物体可能是什么
to help predict probabilities for what theobject is,
根据高水平的过滤去激活
based on the high level filter activations.
我答对了
And I got it right!
对于这些测试很棒的一点是
What ’ s great about these quizzes is
题目彼此是有关联的
that they keep building on each other,
所以即使我这道题对了
so even though I got that one right,
我还需要得到更多的信息
I ’ m still getting more information throughout and
每个问题都会更有趣一点
each question gets a little bit more interesting.
如果你错了一道 也可以
And if you get one wrong, that’s ok too!
因为重点不是要考好
Because the point isn ’ t to beat the quiz,
而是不断学习
it ’ s to keep learning,
就像这些神经网络一样
just like these neural networks do! So,
如果你想要检测一下自己的神经网络知识
if you want to test out YOUR neural network,
前200名在brilliant.org/scishow报名
the first 200 viewers to sign up
的观众会得到20%的全年优质会员优惠
at brilliant.org/scishow will get 20 % off
同时还资助了科学秀——
their annual premium subscription, and you ’ ll help support SciShow –
在此感谢大家
so thanks!

发表评论

译制信息
视频概述

简要介绍了自动驾驶汽车目前发展状况以及存在的难题

听录译者

收集自网络

翻译译者

greenhoodie

审核员

审核员 AF

视频来源

https://www.youtube.com/watch?v=2XQPTtTv_O4

相关推荐