Humanoid robots are heading into mass production in 2025-2026, with Tesla and Figure AI leading and B-end industrial applications accelerating. Based on the production plans of the head companies — Tesla’s Optimus targeting 2026 volume manufacturing, Figure AI, and domestic players like UBTECH — the 2025-2026 window is the industry’s critical period for scale validation and commercial acceleration.
The sector currently shows “hardware cost reduction” and “software intelligence gains” advancing in tandem, driven first by B-end industrial scenarios, especially automotive manufacturing. UBTECH’s Walker S1 and XPeng’s Iron have already entered the factories of BYD, Zeekr, and XPeng for real production-line training, validating their on-line value. The global competitive landscape has taken shape with divergent paths: overseas giants (Tesla, Figure AI) hold the technical high ground through AI algorithms, systems integration, and end-to-end models, while domestic players leverage China’s automotive and 3C supply chains for supply-chain synergy, rapid iteration, and cost control — with some domestic products already priced down to the 100,000-yuan level, a gap that sets up mass-market penetration.
The Industry Chain Structure
The humanoid chain divides into three tiers: upstream core components, mid-stream body manufacturing, and downstream scenario applications. The mid-stream body makers are the chain’s “OEMs” — occupying the dominant position through technology integration, product definition, and scaled production. Upstream, value concentrates in three component families. Joints are the basis of all motion, in rotary and linear types. Reducers secure motion precision — mainstream options are harmonic reducers (Tesla Optimus’s choice) and planetary reducers, with Japan’s Harmonic Drive and domestic Leaderdrive as representative players. Sensors give robots perception: force/torque, tactile, and vision. Mid-stream is a accelerating technical race: overseas, Tesla (Optimus) and Figure AI (Figure 01) lead the frontier; domestically, startups UBTECH (Walker S1) and Zhiyuan (A1) have emerged alongside cross-over entrants — XPeng (PX5), Xiaomi (CyberOne). Downstream, commercialization focuses on B-end: industrial manufacturing first — especially automotive plants, where Walker S1 trains on BYD and Dongfeng Liuzhou Motor lines — then logistics (carrying and sorting), with penetration into commercial services and homes to follow.
The Architecture and the Six Challenges
Humanoids universally run a “brain-cerebellum-limb” full-stack architecture: the brain handles high-level task planning and decision intelligence; the cerebellum owns real-time motion control and balance; the limbs execute against the physical world. The layered decoupling balances complex-task generalization against high-frequency real-time control — the mainstream technical path. Industrialization now depends on synchronized hardware and software breakthroughs: hardware faces cost, mass production, and battery life; software faces intelligent generalization, data scarcity, and real-time constraints.
On hardware: first, high cost and missing standardization — hardware designs have not converged, BOM costs stay high, and joint modules plus dexterous hands are the cost-reduction core. Second, insufficient mass-production capacity — head companies managed only hundreds-to-thousands of units in 2025, mostly for data collection rather than commerce, though Fourier Intelligence’s harmonic reducers entering mass testing and Unitree’s self-developed M107 joint motor point to a breaking bottleneck. Third, battery life — mainstream products fall short of all-day operation, though GAC’s GoMate with solid-state batteries reaches six hours and Pudu’s D7 exceeds 1kWh for 8-plus-hour shifts. On software: first, insufficient generalization — the field has not had its ChatGPT moment; model representation and high-quality data remain the core bottlenecks, though Zhiyuan’s GO-1 (ViLLA architecture) lifted average task success 32 percent and Figure’s Helix VLA balances generalization and real-time control through a slow-system-plus-fast-system design. Second, data scarcity — motion and manipulation data are complex-modal, expensive to collect, and lack standards; NVIDIA’s Cosmos platform generates physical synthetic data from video world models, and Galbot’s GraspVLA pretrains on synthetic big data. Third, real-time and compute constraints — current model action frequencies (pi0’s 50Hz) fall short of complex-scene demands (target: 100Hz). The intuition: action frequency is how many times per second the robot can “refresh” its motion. Fifty frames a second sounds fast, but in the moment a falling robot must slam a hand to the floor, one missed frame breaks the wrist. From RT-1 to Helix, model control frequency climbed from under 10Hz to 200Hz — only a cerebellum with a high enough refresh rate handles genuine surprises.
Market Size and the Scenario Path
The industry stands at the edge of commercial explosion, with 2025 broadly called the “Year One of mass production.” Long-term, humanoids aim at C-end ubiquity: global shipments projected beyond 70 million units and market size past 10 trillion yuan. Industrialization follows a scenario gradient — industrial manufacturing → commercial services → home service — with factories as the breakthrough beachhead (flexible automotive production lines), commercial services expanding through unmanned retail and office scenarios, and homes remaining an early-stage exploration given complexity and cost.
The Four Camps of Global Players
Overseas giants lead: Tesla’s Optimus is the emblem of the end-to-end model route, built on its AI and autonomous-driving accumulation; Figure AI takes a layered-decision approach — GPT-4V-class visual reasoning on top, high-frequency control beneath — with excellent engineering results. The domestic tiers are stratifying: the first tier — UBTECH, Zhiyuan, Unitree — all valued above 10 billion RMB with strong fundraising and integration power (Walker S1 training in BYD and Dongfeng plants; Unitree’s 700 million RMB C-plus round and the 99,000-yuan Unitree G1 pushing into the consumer market). The second and third tiers — Leju, Pudu, Fourier, Galbot — differentiate through government and anchor-client resources (Huawei, Meituan), open platforms, or core components. Cross-over automakers bring manufacturing discipline: XPeng’s Iron robots train on the P7-plus assembly line in its Guangzhou plant with L3-level mass production planned for 2026 at a cost target under 120,000 yuan; GAC’s GoMate emphasizes fully self-developed core components and automotive supply-chain reuse; BYD and Chery participate through direct investment (BYD in Zhiyuan) and joint ventures. Tech giants play enabler and investor: Huawei’s Pangu model powers Leju, Baidu’s ERNIE partners with UBTECH on task planning, Xiaomi’s self-developed Xiaomi Brain supplies a brain; Tencent has invested in Zhiyuan and Unitree, Alibaba in Unitree and LimX Dynamics; and scenario owners like Meituan invest from business need — Meituan in Unitree and Gaussian Robotics for instant delivery.
Chain One: The Figure Supply Chain
After its September 2025 Series C, Figure’s valuation soared to $39 billion — nearly 15 times its $2.6 billion valuation of February 2024 — with NVIDIA, Microsoft, and Amazon among the backers, marking the track’s shift from concept to the eve of industrial explosion. Value currently concentrates in upstream core components, which together exceed 70 percent of hardware cost and carry the highest technical barriers — the short-term focal point for supply-chain watchers.
Chain Two: The Tesla Optimus Supply Chain
Tesla is accelerating Optimus iteration, with the V3 design in its final finalization stage — the goal being technical freeze to clear the road for scale manufacturing. The new generation focuses on two breakthroughs: first, a hand as dexterous as a human’s — each arm will integrate 26 actuators, sharply upgrading manipulation precision and coordination; second, deep integration of the Grok voice model, letting the robot understand complex natural-language instructions and plan actions autonomously. Tesla has set a deliberately aggressive capacity-ramp target, aiming to reuse its automotive mass-manufacturing experience to reach million-unit annual capacity. For supply-chain researchers, the two chains together define the industry’s near-term value map: whoever supplies Figure’s precision components and Tesla’s actuator-and-fabrication volume sits at the intersection of the sector’s two scaling curves — and both curves point up at once.
