Humanoid robots are general-purpose bipedal machines being pushed from research demos into paid work by a collapsing Chinese cost curve, but as of mid-2026 the whole global market is roughly $2B of revenue against OEM valuations of tens of billions, and no maker reports $100M of recurring product revenue ref ref.
Summary
A humanoid robot is a general-purpose mobile manipulator built in roughly human form: two parts (or a wheeled base), two arms, dexterous hands, and a head carrying cameras and microphones. The design bet is that human-shaped machines can be dropped into environments built for humans (warehouses, factories, shops, receptions) without re-engineering the environment, and that one hardware platform can be retasked by software rather than rebuilt per job. That is the opposite of the classical industrial robot bet, where a fixed arm in a caged cell is cheap and reliable because the world around it is redesigned to suit it.
The machine decomposes into three cost-and-difficulty layers. First, actuation: electric motors plus reducers (harmonic or strain-wave drives) at each joint. Actuation, including dexterous hands, is the overwhelming majority of a humanoid’s bill of materials, with semiconductors a small and shrinking fraction, and even the on-board silicon mostly motor control rather than AI compute ref. Actuators, reducers, force sensors and tactile systems are the components with the highest BOM share and the steepest qualification barriers ref. Second, sensing: vision, depth, and touch. Tactile stacks are still de facto rather than standardised, with Meta’s Sparsh-X touch encoders trained on around 1M Digit 360 sensor interactions sitting at the integration layer ref, while the IEEE Humanoid Study Group’s published framework addresses general humanoid safety rather than tactile interfaces ref. Third, autonomy: the policy that turns instructions into useful multi-hour work. This is the weakest layer, and much current deployment is teleoperated or teleop-assisted.
The parameters that decide the outcome are unit price, hours of useful work per intervention, and effective labour cost. Prices are falling fast: Unitree’s platforms went from H1 at $90,000 to G1 at $16,000 to R1 at $5,900 in July 2025 ref, average humanoid pricing fell from $50K+ to roughly $27,300 in 12 to 18 months at an estimated ~67% gross margin against a ~$9K G1 bill of materials ref, and IDTechEx forecasts a 68% average price decline by 2030, from $114,700 to about $37,000 ref. Teleoperated deployments are approaching sub-$30/hour effective labour cost ref. Volumes are still small in absolute terms: global shipments are projected to exceed 50,000 units in 2026, with Unitree and AgiBot taking nearly 80% of them ref.
Beside the industrial line of work sits a smaller social/service line: expressive humanoid heads and companion robots. A recent design integrates a 21-DoF articulated head with silicone skin, SCRFD/ArcFace/ByteTrack face recognition and Llama plus Whisper for conversation in a receptionist role, and NSF is funding work on robots that detect emotional distress and respond with speech, gesture and touch, motivated by six-to-eight-month waits for US mental health providers. These share the form factor but almost none of the economics.
Evidence base
- 8 Jan 2024: Goldman Sachs puts humanoid-only revenue at ~$38B by 2035 on >1.4M unit shipments, revised up ~6x from a prior ~$6B estimate 2024 Goldman Humanoid Robot Tam 38Bn.
- 28 Oct 2025: Unitree launched the R1 humanoid at $5,900 in July 2025, following G1 at $16,000 and H1 at $90,000 ref.
- 31 Mar 2026: Ubtech reported 54.6% gross margin on its humanoid business for 2025, with management guiding 40-43% for 2026 ref.
- 9 Apr 2026: TrendForce projects global humanoid shipments above 50,000 units in 2026, with Unitree and AgiBot taking nearly 80% ref.
- 8 Jun 2026: Humanoid pricing fell from $50K+ to ~$27.3K in 12-18 months at ~67% gross margin on a ~$9K G1 BOM, with teleoperated deployments nearing sub-$30/hour effective labour cost and ~10,000 Unitree units shipped by mid-2026 ref.
- 24 Jun 2026: Morgan Stanley raised its China humanoid forecast to 50,000 units and $2B total market revenue for 2026, distributed across dozens of OEMs with none reporting >$100M recurring revenue ref.
- 19 Jul 2026: $6.2B raised by humanoid companies over 12 months, top three deals 47% of the total, with Figure at $39B and Skild at $14B on minimal revenue ref.
Open questions
- What fraction of paid humanoid hours in 2027 are teleoperated or teleop-assisted rather than autonomous? This is the single number that separates a labour-substitution market from a staffing business.
- Does Agility’s $300M+ of contracted bookings convert into reported recurring revenue, and at what gross margin, following the July 2026 SPAC ref?
- Can any non-Chinese actuator or reducer supplier win volume design-wins at scale, or do harmonic and strain-wave drives plus Schaeffler-class incumbents keep the socket ref?
- Does Ubtech’s 54.6% humanoid gross margin hold near its 40-43% 2026 guidance as Unitree’s ASP collapse propagates, or does it compress further ref ref?
Assessment drafted 2026-08-31 from up to 18 KB sources using the technology-scorecard framework; scores are a draft read pending review.