Table of Contents

Boston Dynamics’ Atlas backflip video amazed the world in 2017. What few people knew: that backflip used hydraulic actuators — extremely high power consumption, complex maintenance, essentially impossible to mass-produce. Seven years later, humanoid robots are genuinely moving toward mass production, but the bottleneck has shifted from “can it move?” to a completely different set of problems.

This post dissects humanoid robot hardware — from actuator selection to supply chain realities — to answer why “doing a backflip” and “working reliably in a factory for 10,000 hours” are problems of entirely different difficulty.

TL;DR

The mass production bottleneck for humanoid robots isn’t AI — it’s hardware: high-precision harmonic drives are in short supply, dexterous hand tactile sensors aren’t mature, and battery energy density limits working hours. The leading manufacturers (Figure, Tesla Optimus, Unitree H1) have chosen different technical paths, and no single design will dominate in the near term.

Design Philosophy: The Performance-Manufacturability-Cost Triangle

Humanoid robot hardware design doesn’t optimize a single objective — it navigates three conflicting ones:

Performance: high torque density, fast response, precise control — pointing toward hydraulics or high-torque motors.

Manufacturability: standardized parts, simplified assembly, high reliability — pointing toward electric drive, avoiding hydraulic plumbing complexity.

Cost: reducing per-joint manufacturing cost — pointing toward standard components rather than custom parts.

The current mainstream approach is electric motor + harmonic drive, trading hydraulic power density for manufacturability and reasonable cost.

Core Subsystem Breakdown

Actuators: The Most Expensive Parts

Actuators convert electrical energy into joint motion. Each joint in a humanoid robot needs one or more actuators.

Harmonic drives are currently the most common choice. They provide high reduction ratios (100:1 is typical) in a small package, turning low-torque motor output into the high torque joints need. The problems:

  • Global supply of high-precision harmonic drives is heavily concentrated — key suppliers are Japan’s Harmonic Drive and Nabtesco, plus a handful of Chinese manufacturers
  • Expensive: a high-precision harmonic drive can cost hundreds to thousands of dollars
  • A humanoid robot has 20–40 joints, so this cost directly determines the BOM

Quasi-direct drive motors — promoted by MIT’s Cheetah and subsequent robots — are another direction: low reduction ratios, high backdrivability, letting the robot sense external force inputs (force control). They sacrifice torque density for more natural dynamic behavior.

Sensors: Letting the Robot “Feel” the World

graph LR
    A[Camera RGB/Depth] --> D[Perception System]
    B[IMU Inertial Measurement] --> D
    C[Joint Torque Sensors] --> D
    D --> E[Control Decision]
    E --> F[Actuators]
    G[Tactile Sensors Dexterous Hand] --> D

Proprioception (joint position, velocity, torque) is the foundation of stable walking control. High-precision joint encoders are necessary, but long-term reliability in high-vibration environments is an engineering challenge.

Dexterous hand tactile sensing is currently the weakest link. A human hand has over 17,000 tactile receptors. Existing robot finger tactile sensors are either too low-resolution, too fragile, or too expensive. “Catching a falling leaf” requires sensing extremely light forces — technically far harder than a backflip.

Structural Materials: Weight vs. Rigidity

Humanoid robots need to balance self-weight against load capacity:

MaterialAdvantagesDisadvantagesApplication
Carbon fiber compositeLightweight, high stiffnessHard to machine, expensive, hard to mass produceLimb structural parts
Aluminum alloyMature processing, reasonable costHeavierMain structural components
Titanium alloyHigh strength-to-weight ratioVery expensiveHigh-stress joints
Engineering plasticsLightest, easy to mass produceLow strengthShells, non-load-bearing parts

The Real Mass Production Bottlenecks

Supply Chain Concentration Risk

High-precision harmonic drives, high torque density motors, 6-axis force/torque sensors — these key components come from highly concentrated supplier bases.

In 2024, as Figure, Physical Intelligence, Unitree, and Zhiyuan all moved toward mass production simultaneously, harmonic drive lead times began extending. This is a “everyone competing for the same suppliers at the same time” problem.

Assembly and Quality Consistency

Humanoid robot assembly is far more complex than automotive: joints need precise clearance control, cable management in high-DOF structures is a nightmare, and sealing and waterproofing need to hold up under repeated bending. Boston Dynamics spent over a decade building its manufacturing capabilities.

Battery and Energy Density

Current humanoid robots typically run 1–3 hours per charge. Lithium battery energy density limits how much energy can be stored within a given space and weight budget. If solid-state batteries commercialize between 2027–2030, the impact on robotics could be as significant as AI model improvements.

Overall

Humanoid robot hardware engineering is going through a critical inflection: from “can it work?” to “can it be mass-produced?” This transition requires not just better AI, but an entire hardware supply chain — actuators, sensors, materials — all reaching production quality simultaneously.

For engineers, the challenges and opportunities in this space are very concrete. If you’re working on sensor fusion, joint control, or supply chain engineering, humanoid robotics is a domain where both technical requirements and market scale are growing rapidly.

References

Ask this article

Answers come from this article only. Click any prompt below or open the chat at the bottom right.

🇺🇸 English

In 2017, Boston Dynamics dropped that Atlas backflip video and the whole internet lost its mind. A humanoid robot, launching itself into the air, sticking the landing. It looked like the future had arrived. But here's what almost nobody talked about at the time: that backflip ran on hydraulic actuators. Enormous power draw, brutal maintenance, and essentially impossible to mass-produce. It was a spectacular science project, not a product.

Fast forward seven years, and humanoid robots are genuinely heading toward mass production. But the hard question has completely changed. It's no longer "can it move?" It's "can it work reliably for ten thousand hours on a factory floor?" And it turns out, doing a backflip and quietly, dependably catching a falling leaf — those are problems of wildly different difficulty.

So let's get into the hardware, because that's where the real story is.

Start with the big picture. Designing a humanoid robot isn't about optimizing one thing — it's about juggling three goals that actively fight each other. First, performance: you want high torque, fast response, precise control. That pushes you toward hydraulics or beefy high-torque motors. Second, manufacturability: you want standardized parts, simple assembly, high reliability. That pushes you toward electric drive and away from all that messy hydraulic plumbing. And third, cost: you want to drive down the price of every single joint, which pushes you toward off-the-shelf components instead of custom parts.

You can't max out all three. So the industry has largely settled on a compromise: electric motors paired with harmonic drives. You give up the raw power density of hydraulics, and in exchange you get something you can actually build at scale for a reasonable price.

Now let's break down the subsystems, starting with the most expensive parts in the whole machine: the actuators. These are what turn electricity into motion, and every joint needs at least one. A humanoid robot has somewhere between twenty and forty joints — so this is where your budget goes to die.

The workhorse right now is the harmonic drive. Think of it as a gearbox that takes a fast, weak motor and turns it into slow, powerful joint movement — a reduction ratio of around a hundred to one, packed into a tiny space. Brilliant piece of engineering. But it comes with problems. The global supply is dangerously concentrated — you're basically talking about Japan's Harmonic Drive and Nabtesco, plus a handful of Chinese manufacturers. They're expensive, anywhere from hundreds to thousands of dollars each. Multiply that by dozens of joints, and the harmonic drive alone dictates your entire bill of materials.

There's an alternative philosophy worth knowing about: quasi-direct drive motors, the approach MIT pioneered with its Cheetah robots. Instead of a huge reduction ratio, you use a small one. The tradeoff is you lose some torque density, but you gain something called backdrivability — the robot can actually feel forces pushing back on it, which gives you far more natural, responsive movement. It's a different bet on what matters.

Next, sensors — this is how the robot feels the world. Picture the flow: cameras, an inertial measurement unit, joint torque sensors, and tactile sensors in the hands all feed into a perception system. That system feeds a control decision, and the decision drives the actuators. It's a loop — sense, decide, act, repeat.

There are two layers here. The first is proprioception — the robot's sense of its own body: where each joint is, how fast it's moving, how much torque it's carrying. This is the foundation of not falling over. It demands high-precision encoders that have to survive years of vibration without drifting, which is genuinely hard.

The second layer is where things get humbling: tactile sensing in the hands. This is the weakest link in the entire field right now. A single human hand has over seventeen thousand touch receptors. The best robot fingertip sensors we've got are either too coarse, too fragile, or too expensive. And this is exactly why catching a falling leaf is harder than a backflip. A leaf weighs almost nothing — to catch it, you need to sense an incredibly delicate force. The backflip is about raw power. The leaf is about exquisite sensitivity, and we're just not there yet.

Quick word on materials, because it's a balancing act between weight and stiffness. Carbon fiber is light and rigid but a nightmare to machine and hard to mass-produce, so it shows up in limbs. Aluminum alloy is the reliable middle child — mature, affordable, a bit heavy — so it forms the main structure. Titanium is incredibly strong for its weight but eye-wateringly expensive, so it's reserved for the highest-stress joints. And engineering plastics are the lightest and easiest to produce but weak, so they're for shells and parts that don't carry load.

Okay — now the part that actually decides whether these robots ship. The real mass production bottlenecks.

First, supply chain concentration. Those harmonic drives, the high torque density motors, the six-axis force sensors — they all come from a tiny pool of suppliers. And in 2024, something predictable happened: Figure, Physical Intelligence, Unitree, and Zhiyuan all pushed toward mass production at the same moment. Lead times on harmonic drives started stretching out. It's the classic problem of everyone racing for the same scarce parts at the same time.

Second, assembly and consistency. Building a humanoid is far harder than building a car. Joints need precise clearances. Cable management inside a body with that many degrees of freedom is an absolute nightmare. And the sealing has to survive endless bending and flexing. This is craft, and it's why Boston Dynamics spent more than a decade just learning how to manufacture these things.

And third, batteries. Today's humanoids run maybe one to three hours on a charge. Lithium energy density is the ceiling — there's only so much power you can pack into a given weight and space. But keep an eye on solid-state batteries. If those commercialize somewhere between 2027 and 2030, the impact on robotics could rival what better AI models have done. Suddenly your robot works a full shift instead of a coffee break.

So where does that leave us? The field is crossing a critical line — moving from "can it work?" to "can we build ten thousand of them that all work?" And that jump doesn't need smarter AI. It needs an entire hardware supply chain — actuators, sensors, materials — all hitting production quality at the same time.

Let me leave you with three things to hold onto. One: the bottleneck for humanoid robots isn't the brain, it's the body — the hardware and the supply chain, not the AI. Two: the flashy demos measure power, but the real frontier is delicacy — that falling leaf is harder than the backflip, and tactile sensing is the field's weakest link. And three: watch the boring stuff — harmonic drive supply, assembly craftsmanship, and battery density. Those unglamorous constraints are what actually decide which robots make it out of the lab and onto the floor. The backflip was the easy part. Everything after it is the real engineering.

🇹🇼 中文

從轉手絹、扭秧歌,到高難度空翻、武術,再到半程馬拉松裡跑出打破人類世界紀錄的成績——這兩年,人形機器人的「本體」進化得快得嚇人。你可能也看過那些影片,機器人幾乎每天都在承受車禍等級的衝擊,還動不動被人類圍著推、圍著踹,它到底是靠什麼撐下來的?

《硅谷101》的主持人陳茜就帶著一個看似簡單的問題,訪談了幾位業內人士:造機器人的門檻,真的不高嗎?公司的護城河,又到底在哪?今天這集,我們就跟著那場訪談,把機器人的身體一層一層拆開來看。

先給你一個地圖。機器人身上的硬體很雜,但可以粗分成四大系統:撐起整個架構的「骨架」、驅動骨架運動的「關節」也就是執行器、感知環境跟自身狀態的「感測器」,還有指揮全身的中樞——「電氣與計算」。我們就從骨架開始,一路拆到晶片。

先講骨架。一輛車以時速六十公里撞上假人,那個衝擊力會把假人撞得七零八落。而對人形機器人來說,承受這種力道已經是日常。受訪者說,機器人每一次空翻觸地,身上承受的加速度可能比汽車、比航太都還高,他們實測過,跟汽車撞牆差不多,有幾十個 g。

這就給材料出了一道矛盾的題目:要翻得起來,身體得夠輕;要扛得住衝擊,強度又得夠大,不然一個空翻零件就飛出去了。材料的演進也正是沿著這條線走。世界上第一台全尺寸機器人 WABOT-1 用的是鋼,體重大概一百六十公斤,跳一下就能把地板砸個坑,更別提翻跟頭。後來從本田的 ASIMO、波士頓動力早期的液壓版 Atlas,到第一代特斯拉 Optimus,鋁合金變成主流,密度只有鋼的三分之一。到現在,業界開始摸索鎂合金,密度比鋁又低三分之一;而在膝蓋、腳踝這種經常挨衝擊的部位,就局部換上強度更高的鈦合金。

有意思的是,這些硬骨架幫機器人扛下所有衝擊,供應商賺的卻只是「辛苦費」。受訪者講得很白:骨架賣出去,扣掉金屬本身、扣掉切削掉的廢料,加工費佔比其實低得可憐。說到底就是賣金屬加加工費,沒什麼太大門檻,一旦量起來,加工費會被壓到非常低的水準。

接下來是重頭戲——關節。真正讓機器人做出高難度動作的,就是它,而這也是整個硬體裡故事最多、成本最高、技術最密集的一塊。幾年前看 Atlas 後空翻覺得驚為天人,現在習以為常,背後的關鍵是關節從液壓系統轉向了馬達。受訪者說,以前做不出這麼好的關節,性能都很差,其實很難翻得起來;就是最近這一兩年,關節技術突然突飛猛進。

關節在業內叫執行器,分成旋轉的跟直線的兩種。拿肩膀當例子,它有三個自由度——前後擺、上下抬、內外旋,本質上都是旋轉,靠三個旋轉執行器一組合,手臂就能朝三個方向自由活動;膝蓋通常只需要一個自由度,一個執行器就夠。這兩種執行器裡都有一套伺服系統,由馬達、編碼器、驅動器、感測器組成。最大的差別在於:旋轉執行器是伺服馬達加減速器,直線執行器是伺服馬達加絲槓。

為什麼要減速器?因為馬達天生就是高轉速、低扭矩,轉速輕鬆上萬轉,但輸出的力很小。而機器人的關節,需要在只轉幾度的同時搬動重物,所以得靠減速器把轉速降下來、把扭矩提上去。減速比越大,扭矩越高。業界最常用三種。行星減速器結構小、成本低,但減速比小、扭矩低,多用在手部。諧波減速器減速比大、扭矩強、精度又高,但它是柔性結構,抗衝擊比較差,用在肘、肩這種地方。最後是 RV 減速器,剛性好、抗衝擊強,就放在髖、膝、腰這些要硬扛衝擊的部位。前面說後空翻的力道相當於汽車撞擊,這就解釋了為什麼諧波減速器撐不住那些部位,得換上剛性更好的 RV。

減速器是整個關節裡最難的一環。受訪者形容它是個「不可能三角」:你要大批量製造,齒輪精度跟長時間運行的穩定性要求都非常高。用了一千個小時就開始各種異響、性能下滑,機器人走路就沒以前好,甚至慢慢走歪;做極限動作又可能把裡面的小齒輪直接撞壞。所以他有句話說得很好——造一個減速器不難,難的是造出一萬個性能一致、又耐用、摔倒還扛得住衝擊的減速器。

再看直線執行器。它其實是最像人體肌肉的——你手臂擺動,不是關節主動在轉,而是連接兩端骨頭的肌肉在收縮。它只做推拉,有些機器人的膝關節就用它模仿膝部肌肉;多個一組合,還能做出手腕、腳踝的旋轉。最簡單的直線執行器就是液壓,老版 Atlas 主要就用直線液壓缸,爆發力強、抗衝擊、功率密度大;但液壓系統複雜、容易漏油、精度也不如馬達,所以新版 Atlas 也轉向馬達了。馬達只能轉,要輸出直線運動就得靠絲槓,原理就像拧螺絲。為了減摩擦,內部會加滾珠,或換成滾柱,壽命更長、剛性更好。不過整體來看,直線執行器在業界用得比較少,就三個原因:動態性能差、製造難、成本高。它載荷可以做大、不供電還能自鎖保持姿勢,但也因為載荷大、減速比大,動作沒那麼敏捷,又很難低成本量產。所以目前量產最多的,還是旋轉關節。

講到馬達本身,它的難點集中在三塊。第一是散熱。一旦過熱,控制系統只能降功率,機器人可能空翻到一半突然「腿軟」倒地。受訪者說最早期的樣品,這種極限動作十分鐘只能做一次,因為做完性能曲線整個就變了,得先冷卻。關鍵是發熱不是線性的:關節做一次極限動作,瞬時電流可能是平常的三到五倍,發熱量就是額定的九到二十五倍,一次空翻,關節溫升可能從十度直接跳到五十度。所以能效百分之三跟百分之五之間,差異巨大,直接決定你敢不敢把性能往上抬。第二是體積,越大轉動慣量越大,改變運動狀態越難——就像甩繩子,繩子越長轉得越慢、要停下來也越費勁。第三是性能穩定性,也就是 TN 曲線,講的是你輸入多少電流,能得到多少轉速跟扭矩。走不平路面時,腳踝的力矩感測器感知起伏,要動態調整電流;如果曲線不穩,同樣指令輸出的扭矩卻有偏差,結果就是摔。而且這條曲線還影響演算法訓練——機器人先在模擬裡練,如果模擬跟現實差太多,一到現實就走樣。

那執行器到底有多貴?據美國銀行測算,它是機器人身上成本最高的部件,大概佔百分之五十一。受訪者打了個比方:馬達加控制器,比骨骼、比眼睛、比腦子、比心臟都貴,所以它就是未來量產降本的關鍵。最主要的推力,是中國供應鏈太捲了——做馬達的臥龍電驅、做減速器的綠的諧波、雙環傳動,甚至能直接提供整套執行器的三花智控、拓普,全都有。

那既然市面上買得到現成的,機器人公司為什麼還要自己苦哈哈研發?兩條路各有取捨。買成品,研發成本低、開發快,但物料貴、難客製、性能也將就——人家賣的是標準件,不會專門為你設計。自研呢,能更好地匹配需求跟演算法、性能更強,但要付出大量研發精力。受訪者的調查是,目前頭部公司更偏向自研,甚至會直接進駐供應商那邊參與設計。因為關節從來不是把零件裝一起這麼簡單,而是要在極小的體積裡,同時平衡力量、精度、耐久、成本、重量——這是整個身體最難的地方。

拆完關節,來看感測器。不管人類怎麼圍毆,機器人大多很難摔,靠的就是一整套感測器。本體感知,靠關節裡的編碼器跟力矩感測器,以每秒上千次的頻率感知每個關節的位置跟受力。IMU 慣性測量單元,相當於人的內耳前庭,感知身體的傾斜跟旋轉,核心是加速度計跟陀螺儀。你被踹一腳,身體瞬間獲得加速度、往旁邊倒,IMU 偵測到,把資料傳給「小腦」,算出各關節該增減多少扭矩,把身體拉回來。這類東西手機、汽車上用得多,相對成熟。視覺系統呢,常見是攝影機加光達加毫米波雷達的多感測器融合,跟汽車自駕很像;例外是特斯拉,馬斯克是堅定的純視覺派,只用攝影機。

不過要注意,同一種感測器,機器人的規格跟汽車差很多。拿比較貴的光達來說:汽車要跑高速,得看到一百五到兩百公尺外;機器人在室內,十到二十公尺就夠,測距短,功率、體積、成本全都能更低。汽車辨識的是車、人、路障這種大物體,點雲可以稀一點;機器人要在桌上拿螺絲起子、在地上撿硬幣,需要密得多的點雲。可靠性上,汽車常年在室外,要撐零下四十度到八十五度;機器人現階段不用這麼寬,但受的衝擊大得多——汽車出車禍的加速度,才勉強到機器人日常空翻一次的水準。所以汽車光達很多為可靠性做的設計,在機器人看來是冗餘的。也因此,雖然汽車光達已經很成熟,機器人光達還在非常早期。

再說觸覺,這塊特別值得講。它主要有四種途徑:壓阻式、電容式、壓電式,還有目前最熱門的光學式——表面彈性材料受力形變,再用攝影機捕捉。理想的觸覺最好是三維的,不只感受壓力,還能感受平面上的摩擦。比如你拿一罐飲料往上抬,手指一感受到向下滑的摩擦力,就會自動加大捏的力度,避免它掉。但這對材料跟演算法都是大挑戰:任何材料都很難在三個方向上乾淨地解耦。所以你會發現,二〇二五年量產的產品裡幾乎不搭觸覺——整個行業都用得很少,因為不穩定:長期抓東西稍微一變形,輸出信號可能就完全不一樣。到二〇二六年才有點變化,受訪嘉賓說看到了規模化生產的希望。但整體而言,觸覺這行還非常早期。

最後是電氣與計算,也就是大腦跟小腦的分工。業界在演算法上發展出 System 1 加 System 2 的雙系統,晶片上也對應「小腦加大腦」。為什麼不用一顆晶片搞定?因為需求剛好相反。大腦晶片思考「該怎麼做事」,要高算力、大記憶體,最好能在端側跑大模型,延遲個幾秒無所謂。目前絕大多數機器人的大腦選 NVIDIA 的 Orin,而二〇二五年 NVIDIA 推出專為機器人設計、性能更高的 Thor,預計會成為未來主流。特斯拉又是例外,用自研晶片,而且是雙晶片。小腦晶片剛好相反,不需要多高算力,但即時性、穩定性、響應速度必須拉滿,延遲個幾毫秒就可能摔。機器人空翻、跳舞基本是提前錄好的動作,但它腳下那些細碎的踏步,就是小腦在動態調平衡,像人的本能反應,頻率可能高達一千赫茲。小腦晶片通常是 MCU,主流是意法半導體的 STM32 這類。

現在還有個新趨勢:把大腦跟小腦整合到一顆晶片上。好處是整個胸腔的體積跟走線更簡潔,而且大腦小腦之間的協調越往後越重要——就像有人朝你扔飛鏢,看軌跡、做預測是大腦,伸手去抓是小腦,兩者通訊越快越有利。做在同一顆晶片上,通訊會非常快。不過業內普遍認為,這條路還很早期,得等出貨夠多、市場夠大,機器人公司才會像今天的智慧汽車那樣,逐漸轉向一體化的自研晶片。當然,別忘了還有像心臟一樣供能的電池,跟像神經血管一樣遍布全身的線束。

好,拆到這裡,我們回到最初那個問題。前段時間的機器人半程馬拉松上,有的隨地大小坐,有的跑著跑著腳崴了、手臂掉了、被減速帶絆倒後「粉身碎骨」;但也有表現亮眼的——榮耀的機器人不但包攬前六名,還刷新了人類的半馬紀錄。連手機廠商做機器人都能這麼強,是不是代表這行業沒門檻?業內人士的回答很妙:「Yes and No」。

Yes 的部分是,前面拆的這些零部件,供應商跟手機、汽車行業高度重合,往上,演算法也有一部分能跟自駕複用。這就是榮耀、小米、特斯拉、小鵬會下場的原因。受訪者估算,電氣與計算的供應商重合度能到九成以上,機械結構、感測器也大同小異,整體八成以上的東西是可以同質的。理論上,你只要認識這些供應商,就能自己搓出一台機器人。

但 No 的部分才是關鍵:「能動」跟「好用」之間,隔著一道巨大的鴻溝。組裝後重量分布不均,重心就偏,走路為了平衡,某些關節得額外出力,功耗上升、續航下降,甚至影響步態。你為了散熱或減重做的每一個選擇,都會讓另一端的性能變差——想讓它更輕,它就扛不了多少重量。實驗室裡跑一小時沒事,放到真實環境跑一百小時,毛病全冒出來:某個螺絲鬆了、某根線磨損了、某個關節的潤滑脂乾了、某個感測器開始漂移,這些都得靠不斷調試,才能找到那個平衡點。

所以受訪者最後這句話,我覺得是整篇的題眼:把每個零部件拆給各家供應商,單看難度都不高,最後的系統整合,才是最難的。

那我們就用三個要點來收尾。第一,人形機器人這兩年的躍進,關鍵不在 AI,而在本體硬體,尤其是關節從液壓轉向馬達這一步。第二,執行器是整台機器人裡最貴、也最難的部分,大概佔一半成本,難不在造一個,而在造一萬個性能一致、又耐用的。第三,也是最重要的——真正的護城河,從來不是你認識哪家供應商,而是在人形這個嚴苛的約束下,把力量、精度、耐久、成本、重量,一次又一次調到平衡,而且做到一萬台,台台都一樣。

Tags

Related Articles