2021 年的特斯拉 AI Day 上,一个穿紧身衣的真人套着银色外壳在台上扭动,扮成机器人。这段画面被嘲笑了整整一年,成了"特斯拉画饼"的经典素材。
2022 年同一个舞台,马斯克上台后的第一句话就是这个梗。
去年那台 Optimus 就是一个人穿着机器人套装。但我们已经走了很远,跟那个比,接下来的东西会非常震撼。(马斯克)
然后他把真的那台请上了台。
这台机器人自己走上舞台,朝观众挥手,抬起手臂做了几个动作。整个过程里,它身上没有缆绳,没有龙门架,没有机械支撑,什么都没有。马斯克在它出场前特意强调了一遍:这是第一次。
这是我们第一次尝试让机器人完全没有任何备份支撑,没有起重机,没有机械结构,没有线缆。什么都没有。(马斯克)
这句话的分量,懂行的人才知道。
人形机器人最难的地方,不在让它动,在让它自己站稳。两条腿加一个高重心,本质上是一根随时会倒的倒立摆。绝大多数实验室里的人形机器人做公开演示时,背后都会挂一根安全绳,或者架一个龙门架,防止它摔个狗啃泥。特斯拉这次把这根绳撤了。
马斯克自己也没敢把话说满。
机器人能做的其实比我们刚才展示的多得多,只是我们不想让它当众脸着地。(马斯克)
先看清楚它到底动了什么
现场演示很短,就几步路加几个挥手动作。真正的内容在视频里。
视频里这台代号 Bumble C 的机器人,在一个类似工厂车间的环境里走来走去,用一把洒水壶给花浇水,从台面上抓起一个箱子,还跳了一段舞。旁边配了它的"第一人称视角":画面是对着世界的三维渲染图,物体被框出来并打上标签,机器人知道自己该抓哪一个。
这套视觉系统,特斯拉没有重新做。他们直接搬了车上那套。
它跑的就是你车里那台简单的自动驾驶计算机。我们用的是和 Autopilot 完全相同的一套流程:收集数据、训练网络、部署到机器人上。(特斯拉工程师)
Bumble C 是台开发机,大量使用了半成品的现成执行器,各个关节是拼凑出来的。特斯拉同时展示了一台更接近量产形态的 Optimus:执行器、电池包、控制系统全部由特斯拉自己设计。这台当时还不会走路。
它还没完全准备好走路,但我想几周之内它就能走。我们想给你们看一台真正接近量产版本的机器。(马斯克)
两万美元,比一辆车便宜
马斯克轻描淡写地抛出一个数字。
Optimus 的设计目标是能力极强,但以极高的产量制造,最终可能达到数百万台。它的成本预计远低于一辆汽车,大概低于两万美元。(马斯克)
两万美元是什么概念。波士顿动力的 Atlas 造价在百万美元量级,而且那是实验室里的手工艺品,一台一台手工攒的。日本 ASIMO 当年的单台成本也超过百万美元,最后因为没法量产而停产。
马斯克给 Optimus 定的路子,从头到尾都是汽车工业的逻辑:不追求造一台最强的机器人,追求造一台能下线的机器人。
我认为很少有人真正意识到 Optimus 的潜力有多大。(马斯克)
这句话没有展开,但整场发布会的技术细节都在为它做注脚。两万美元的售价意味着什么?意味着这台机器人的每一个零件、每一道工序、每一次供应链谈判,都得按 Model 3 的标准来做。实验室可以做出一台不会摔的机器人,但做不出一万台每台两万美元的机器人。
电池够干一整天活
负责整机架构的工程师放出了结构图。橙色是执行器,蓝色是电气系统。
躯干正中间装的是电池包,容量 2.3 千瓦时。这个数字听着不大,大致相当于一台家用冰箱运行一天的耗电量,但工程师对它的定位说得很明确:够一整天的活儿。
电池包容量是 2.3 千瓦时,正好够一整天的工作量。它特别的地方在于,所有电池电子部件都集成在电池包内的单块 PCB 上。(特斯拉工程师)
把所有电池管理电路塞进一块板子,好处是体积、线束和成本一起降。这是典型的汽车工程思路:把结构做简单,让它在流水线上好造、好修、好降本,性能堆到极限不是第一优先级。
机器人的"脑子"不在脑袋里
结构图上有个细节值得停一下。中央计算机不在头部,而在躯干。
接着是我们的"大脑",它不在头里,但离得很近,也在躯干部分。(特斯拉工程师)
它的工作内容描述得很直白:处理视觉数据,根据多路传感器输入做出瞬间决策,还负责通信。
通信这块配了无线连接和音频支持。另外还有一项容易被忽略的设计:硬件级安全特性。
它还有硬件层面的安全功能,用来同时保护机器人和它周围的人。(特斯拉工程师)
一台两万美元、会走路、有 28 个关节、能举起半吨重物、还连着无线网络的机器,安全必须是硬件级的,不能只靠软件里写几行条件判断。这是所有做人形机器人的公司都绕不开的坎,特斯拉在这台原型机上就已经把它列进了架构。
28 个执行器,举得起一台九英尺三角钢琴
执行器这部分,信息密度最高。
Optimus 全身有 28 个执行器。工程师先做了一个类比:汽车的动力总成和机器人有大量相似之处,核心都是能量、质量和成本这三件事。但有个根本区别。
汽车有两个驱动单元,用来把车从 0 加速到 60 英里每小时,或者支撑城市路况驾驶。而机器人有 28 个执行器,执行器层面的任务并不明显,我们有更高层的任务,比如走路、爬楼梯、搬重物,这些都需要被翻译成一个个关节的规格。(特斯拉工程师)
这句话点出了人形机器人真正的工程难点。人类走路的时候不会想"我的髋关节该输出多大扭矩",但设计机器人的时候,你必须把"走路"这个模糊的目标,拆成 28 个关节各自的扭矩、转速、精度和重量预算。拆错了,整机就废了。
旋转执行器是重点。特斯拉给的方案是自研的:高速端用角接触球轴承,低速端用交叉滚子轴承,齿轮组是应变波齿轮,也就是谐波减速器,内置三个传感器,电机是定制的永磁电机。另外集成了一套机械离合器,可以在过载时断开保护。
这些名词堆起来不好读,但有一个数字能立刻让人明白它的水平。
我们的执行器能举起半吨重的东西,一台九英尺的音乐会三角钢琴。(特斯拉工程师)
半吨是什么概念。一台九英尺的施坦威音乐会三角钢琴,重量在 400 到 500 公斤之间。这个数字不意味着机器人能把钢琴扛在肩上跑,它衡量的是单个执行器在额定工况下的扭矩输出换算出来的负载量级。28 个关节的力量密度,远远超过了"能挥手"的演示级别。
一只手的 11 个自由度
手部设计的讲解,细致到近乎啰嗦。
一只手的配置是 6 个执行器、11 个自由度。手指由金属肌腱驱动,这种肌腱既要柔韧又要结实。工程师强调了两项能力的同时存在:大开口的力量抓握,以及针对细小、轻薄、易碎物体的精密捏取。
我们有能力完成大开口的力量抓握,同时也针对小、薄、易碎物体的精密抓取做了优化。(特斯拉工程师)
这两个需求在工程上经常打架。要力量,就得粗壮的传动;要精细,就得轻巧的结构。特斯拉的解法是用金属肌腱把执行器移出手掌,手掌本身做轻,力量从手臂传过来。
手掌里放了一个手内控制器,直接驱动手指,同时接收来自 Autopilot 的传感器反馈。也就是说,手指的触觉回路和视觉系统是同一套体系里跑出来的。
量产版 Optimus 的手还有一项设计:拇指有两个自由度,可以做出真正的对掌动作,左右手都有。
它能移动所有手指,拇指有两个自由度,所以它拥有对掌拇指,左右手都是。这样它就能操作工具,做真正有用的事情。(特斯拉工程师)
"操作工具"这四个字,是人形机器人整个命题的起点。人类造的所有工具,从螺丝刀到电钻到方向盘,都是按人类的手设计的。机器人想要接手人类的活儿,先得有一双结构上对得上的手。
视觉系统几乎是原样搬过来的
AI 部分,特斯拉最有底气。
机器人用的占用网络(Occupancy Network),和车上用的是同一个。工程师说得很直接:唯一改变的东西是训练数据,需要重新采集。
这段视频里机器人跑的就是完全相同的占用网络。唯一真正变化的是我们不得不重新收集的训练数据。(特斯拉工程师)
这句话背后是特斯拉投入多年的基础设施:数据引擎、自动标注、大规模训练集群、仿真环境。这套东西一旦建成,换一个载体去用,边际成本极低。别人从头做一个机器人视觉系统要几年,特斯拉换了一批数据。
他们还在用神经辐射场(NeRF)做改进。
我们也在尝试用神经辐射场方面的工作来改进这些占用网络,为机器人所处的环境获得非常好的体积渲染效果。(特斯拉工程师)
NeRF 能从一堆二维照片里重建出三维场景,用很低的采集成本就能获得带纹理、带光照的真实环境模型。这既可以用来做训练数据,也可以用来在仿真里批量生成随机场景。
还有一个技术细节:机器人在移动时,需要知道自己在哪里、朝哪个方向。
我们训练了更多神经网络来识别机器人摄像头画面里的高频特征和关键点,并在机器人导航过程中跨帧追踪它们。我们再用这些点来更准确地估计机器人在环境中的位姿和轨迹。(特斯拉工程师)
这套逻辑和车上的视觉里程计是同一个思路:不依赖昂贵的激光雷达和高精地图,靠摄像头画面里稳定出现的特征点,反推自己走了多远、拐了多少度。
四月的蹒跚学步,到八月的行走自如
运动控制部分放了一段仿真里的进度回放,整场发布会里这段最有说服力。
这是运动控制代码在我们自研仿真器里运行的效果,展示了机器人行走能力的演变过程。你可以看到我们四月份起步非常慢,随着我们解锁更多关节、采用更深入更先进的技术,比如手臂平衡,进度开始加速。(特斯拉工程师)
四月的机器人走起来像刚学步的幼儿,摇摇晃晃,每一步都在找平衡。到了后期,步态明显稳定,手臂开始配合摆动,速度也上来了。
手臂平衡这件事值得单独说一下。人类走路时手臂自然摆动,这不是装饰,是在抵消躯干旋转、帮助维持动态平衡。让机器人的手臂参与平衡控制,意味着控制系统要同时协调腿和手臂的几十个自由度,难度比"两条腿各动各的"高一个量级。仿真里那条曲线在这个阶段明显抬升,说明这不是线性的进步,是解锁了某个能力之后才开始加速。
让机器人学会"像人一样"拿东西
操作能力这块,特斯拉的方法拆成两步。
第一步,是人类动作的采集与迁移。他们让真人做一遍动作,比如拿起一个盒子,用动捕记录下这个过程中手、肘、躯干各个关键点的位置,形成一串关键帧。然后用逆运动学把这套动作映射到机器人身上,形成一个"自然动作参考库"。
但工程师马上指出了问题所在。
单次演示无法泛化到真实世界的各种变化。比如说,这个演示只对放在某个特定位置的盒子有效。(特斯拉工程师)
动作库里存的只是"在那个位置拿起那个盒子"这一条轨迹。盒子挪了 20 厘米,这条轨迹就废了。
第二步解决问题。他们把参考轨迹丢进一个轨迹优化程序,让优化器自己去算:手应该放在哪,身体该怎么配重,动作要怎么在真实世界里调整。
如果盒子在这个位置,优化器就会生成这样一条轨迹。(特斯拉工程师)
这套思路把"像人"和"能用"拆成两件事。人类动捕保证动作自然、符合人类环境的物理直觉;优化器负责把它适配到当下的具体场景。前者是模板,后者是适配层。
这个设计选择很务实。纯强化学习能学到很厉害的动作,但学出来的姿势往往怪异,人类看着不舒服,而且在人类环境里容易撞到东西。纯动捕回放又完全不能泛化。两头各取一半,是当下最可能跑通的路。
接下来几周要干什么
发布会结尾,马斯克给出了时间表,具体到有点不像他平时的风格。
我想接下来几周的第一件事,是让 Optimus 至少达到 Bumble C 的水平,也就是你们刚才看到的那台原型机,可能还要超过它。我们也会开始聚焦工厂里的真实用例,让这个产品成为现实,改变整个经济。(马斯克)
注意他用了"改变整个经济"这种说法。这不是随便说说。如果一台能干体力活的机器人成本是两万美元,能连续工作一整天,能充电,能重复使用多年,它对劳动力市场的影响不按台数计算,按倍数计算。
然后他补了一句,语气像是在给自己打气。
所有这些,都是在短短六到八个月里完成的。(马斯克)
从零到一台能自己走上台、自己保持平衡、用视觉识别物体、用肌腱手抓东西的人形机器人原型,六到八个月。这个速度放在人形机器人行业里,几乎不正常。波士顿动力做了二十多年,ASIMO 做了十几年,而特斯拉的第一版原型机是在 2021 年 8 月才第一次公开亮相的。
当然,原型机和量产机之间隔着一条巨大的鸿沟。这台机器人还不能干真正的活,走路速度慢,演示环境是精心布置的,视频里的操作也是剪辑出来的。马斯克自己都说"我们不想让它当众脸着地"。
但有两件事是可以确认的。
一是这套技术栈的复用效率。视觉来自 Autopilot,训练流程来自 Autopilot,仿真和数据处理的基础设施来自 Autopilot,电池和电机工程的思路来自整车。特斯拉做机器人,走的是把已有能力搬到新载体的路子,而不是另开一条战线。这是它跟所有机器人公司最本质的区别。
二是成本路线。两万美元的定价不是随口定下的目标,它决定了整台机器人的每一个设计决策:把电池电子塞进一块 PCB,把执行器做成可量产的模块,让手部执行器移出手掌以减轻末端重量。所有细节都在朝"能大批量下线"这个方向走。
人形机器人这个赛道,过去几十年一直卡在同一个地方:技术演示很惊艳,商业上不成立。特斯拉这次给出的答案很粗暴,把成本打下来,把产量提上去,把技术从车上搬过来。
这个答案对不对,几周后 Optimus 能不能真的走到 Bumble C 的水平,工厂里的第一份真实工作能不能干下来,才是验证的开始。
发布会最后,马斯克站在台上,身后是两台银色的机器人,一台会走,一台更接近量产但还不会走。他说了句谢谢,灯光就暗了。
一年前那个穿紧身衣的人,这次没有出现。
如果你关心这个行业,可以关注这几件事
- 盯住下一代 Optimus 的行走视频。这次演示是无缆绳的,但速度慢、步态保守。什么时候它能在不平整地面上快走、遇到推搡能自我恢复,运动控制才算真正过关。仿真曲线好看和真机稳定是两回事。
- 盯住工厂试点的具体岗位。马斯克说下一步是把机器人放进自家工厂,那就看它被安排做什么:搬运、分拣、上料,还是只是巡检。任务越接近标准化的重复劳动,商用推进越快;越需要柔性判断,推开越慢。
- 盯住两万美元这个价能不能守住。它决定了整个商业模型成不成立。如果最终价格翻倍到四五万美元,市场空间会立刻缩水一大半。供应链的国产化程度、执行器的良率、电池包的成本,都是关键变量。
- 盯住手的自由度是否保留到量产。11 个自由度、6 个执行器、金属肌腱这套方案成本不低,很多公司在量产阶段会砍掉手指的冗余自由度换取便宜。如果特斯拉守住这个配置,说明它真的打算让机器人操作工具,而不是只做展示。
- 如果你是开发者,现在就可以开始熟悉占用网络、逆运动学和轨迹优化这三块。这套技术栈的复用性很强,从自动驾驶到机器人操作都能用得上。人形机器人的人才需求,未来几年会沿着这条线爆发。
0:00welcome to Tesla AI day 2022. I do want to set some expectations with respect to our Optimus robot as as you know last year it was just a person in a robot suit but We've now we've come a long way and it's I think you know compared to that it's going to be very impressive so should we should we bring out the vot before we do that we have one one little bonus tip for the day this is actually the first time we try this robot without any backup support cranes mechanical mechanisms no cables nothing yeah want to join with you guys tonight but it was the first time let's see you ready let's go [Applause] [Music]
1:19foreign [Music] so this is essentially the simple self-driving computer that runs in your Tesla cars by the way [Music] this is the this is literally the first time the robot has operated without a tether was on stage tonight so the robot can actually do a lot more than we just showed you we just didn't want it to fall on its face so we'll we'll show you some videos now of the robot doing a bunch of other things yeah we wanted to show a little bit more what we've done over the past few months with apart and just walking around and dancing on stage ah just humble beginnings but you can see the autopilot neural networks running as is just retrained for the bud
2:39directly on that on that new platform that's my watering can yeah when you see a rendered view that's that's the robot what's the that's the world the robot sees so it's it's very clearly identifying objects like this is the object it should pick up picking it up we use the same process as we did for autopilot to connect data in train your networks that we then Deploy on the robot that's an example that illustrates the upper body a little bit more what you saw was what we call Bumble C that's our sort of rough development robot using semi-off-the-shelf actuators but we actually have gone a step further than that already the team's done an incredible job
3:22and we actually have an optimist bot with a fully Tesla designed at both actuators battery pack control system everything it it wasn't quite ready to walk but I think it will walk in a few weeks but we wanted to show you the robot the the something that's actually fairly close to what will go into production and and show you all the things it can do so let's bring it out let's do it [Music]
4:14all right [Music] with the degrees of freedom that we expect to have in Optimus production unit one which is the ability to move all the fingers independently move the to have the thumb have two degrees of freedom so it has opposable thumbs and both left and right hand so it's able to operate tools and do useful things our goal is to make a a useful humanoid robot as quickly as possible the Optimus is designed to be an extremely capable robot but made in very high volume probably ultimately millions of units and it is expected to cost much less than a car I would say probably less than twenty thousand dollars the the potential for optimistic is I think appreciated by
5:15very few people hey as usual Tesla demos are coming in hot software integration Hardware upgrades over the months since then but in parallel we've also been designing the Next Generation this one over here obviously there's a lot that's changed since last year but there's a few things that are still the same you'll notice we still have this really detailed focus on the true human form so on the screen here you'll see in Orange are actuators which we'll get to in a little bit and in blue our electrical system so in the middle of our torso actually it is the Torso we have our battery pack the sized at 2.3 kilowatt hours which is perfect for about a full day's worth of work
6:03what's really unique about this battery pack is it has all of the battery Electronics integrated into a single PCB within the pack so going on to sort of our brain it's not in the head but it's pretty close also in our torso we have our Central computer so we still are gonna it's gonna do everything that a human brain does processing Vision data making Split Second decisions based on multiple sensory inputs and also Communications so to support Communications it's equipped with wireless connectivity as well as audio support and then it also has Hardware level security features which are important to protect both the robot and the people around the robot so now that we have our
6:44sort of core we're going to need some limbs on this guy and we'd love to show you a little bit about our actuators and our fully functional hands as well so there are many similarities between a car and the robot when it comes to powertrain design the most important thing that matters here is energy mass and cost in the particular case you see a car with two drive units and the drive units are used in order to accelerate the car 0 to 60 miles per hour time or drive a City Drive site while the robot that has 28
7:18actuators it's not obvious what are the tasks at the actuator level so we have tasks that are higher level like walking or climbing stairs or carrying a heavy object which need to be translated into joint into joint specs the rotary actuator in particular has a mechanical clutch integrated on the high speed side angular contact ball bearing and on the high speed side and on the low speed side a cross roller bearing and the gear train is a strain wave gear and there are three integrated sensors here and the bespoke permanent magnet machine so our actuator is able to lift
8:12a half tone nine foot concert grand piano our fingers are driven by metallic tendons that are both flexible and strong we have the ability to complete wide aperture power grasps while also being optimized for precision gripping of small thin and delicate objects some basic stats about our hand is that has six actuators and 11 degrees of freedom it has an in-hand controller which drives the fingers and receives sensor feedback reported directly from autopilot to the Bots situation it's exactly the same occupancy Network that we're talking to a little bit more details later with the autopilot team that is now running on the butt here in this video the only thing that changed
8:53really is the training data that we had to recollect we're also trying to find ways to improve those occupancy networks using work made on your Radiance fields to get really great volumetric rendering of the Bots environments for example here some Machinery that the bot might have to interact with so we've been training more neural networks to identify high frequency features key points within the Bots camera streams and track them across frame over time as the bot navigates to its its environment and we're using those points to get a better estimate of the Bots pose and trajectory within its environment as it's walking and this is a video of the
9:34motion control code running into your product simulator simulator showing the evolution of the robot's work over time and so as you can see we started quite slowly in April and start accelerating as we unlock more joints and deeper more Advanced Techniques like arms balancing over the past few months we wanted some manipulate objects while looking as natural as possible and also get there quickly so what we've done is we've broken this process down into two steps first is generating a library of natural motion references or we could call them demonstrations and then we've adapted these motion references online to the current real world situation so
10:11let's say we have a human demonstration of picking up an object we can get a motion capture of that demonstration which is visualized right here as a bunch of keyframes representing the locations the hands the elbows the Torso we can map that to the robot using inverse kinematics and if we collect a lot of these now we have a library that we can work with but a single demonstration is not generalizable to the variation in the real world for instance this would only work for a box in a very particular location so what we've also done is run these reference trajectories through a trajectory optimization program which solves for where the hand should be how
10:53the robot should balance during when it needs to adapt the motion to the real world so for instance if the box is in this location then our Optimizer will create this trajectory instead I think the first thing within the next few weeks is to get optimists at least at par with Bumble C the other bug prototype you saw earlier and probably Beyond we're also going to start focusing on the real use case at one of our factories and I make this product a reality and change the entire economy all of this was done in barely six or eight months thank you very much [Applause]
资源与出处
- 视频在 YouTube 收看
- 档案silicon-valley.tech 档案页本页视频与文稿数据来自 Elon Musk Archive
- 精读本篇精读 Markdown 版