Unitree Market Research

1. Introduction

Unitree Robotics is the embodiment vendor with the highest degree of hardware scale, and the only one among the ten vendors in this group to complete an A-share IPO. Its primary narrative is "hardware scale and first stock to list": it closed the commercial loop starting from quadruped robots, then used humanoid robots to complete the leap from the research-and-education market to the capital markets. On the Omdia metric (as quoted in Zheshang Securities, 2026-03), global humanoid robot shipments in 2025 were about 13,000 units, with Unitree ranking world No. 1 at over 5,500 units.

1.1. Company Overview

1.2. Development History and Key Milestones

1.3. Business Model and Pricing

Unitree's business model is standard hardware product sales + ecosystem add-ons: embodiments are sold directly at published list prices by model (mainly B-side research and education, with the C-side launched in recent years), layered with accessories, software, and a developer ecosystem. Two key pricing facts:

1. Per the WRC 2026 observation metric reported by Sina Finance, the unit price of its own products fell from about 590,000 yuan to about 170,000 yuan within two years, and the R1 pushed it further down to 29,900 yuan, actively breaking through the industry's lowest price band.
2. No official figure for the specific H2 pricing was found; marked [To be filled].

The IPO use-of-proceeds disclosure reveals a clear capacity plan: annual capacity of 190,000 units, of which 75,000 are humanoid robots and 115,000 are quadruped robots. This indicates that the foundation of Unitree's business model is still cost amortization through hardware mass production; software and ecosystem (App Store, WVLA models, GR00T collaboration) are the second growth curve rather than the current main body of revenue.

2. Glossary

3. Feature Description and Product Lines

3.1. Product Line Overview

3.2. Core Capabilities

1. Dynamic motion capability: G1 completed the world's first humanoid side flip and 720-degree spinning kick (2025-03); H2 training videos demonstrate highly dynamic actions such as flying kicks and kicking a watermelon apart (2026-01). Such demonstrations are public proof of Unitree's motion control (cerebellum layer) capability.
2. Autonomous tasks by the embodied foundation model: WVLA2.0 deployed to autonomous meeting-room tidying on G1 (item placement, categorized storage), with no teleoperation throughout; G1's full-chain voice interaction was upgraded, responding in real time from natural language through semantic understanding and spatial perception to trajectory planning and action execution.
3. Fleet orchestration: all three Spring Festival Gala performances required fully autonomous multi-robot coordination and positioning, a large-scale public validation of L3 orchestration-layer capability.
4. Developer ecosystem: announced the humanoid robot "App Store" (2025-12), attempting to open the embodiment as a platform-type entry point for skill distribution.

4. Platform Architecture

4.1. Brain–Cerebellum–Embodiment Stack

Unitree's public architecture information is most complete at the embodiment and motion control layers; the foundation model layer has been rapidly filled in over recent years:

4.2. Data Loop

Unitree has not disclosed a self-built data collection factory (contrast: AgiBot's self-built data collection factory collects 30,000–50,000 items daily, see 02-agibot.md for details). The publicly visible parts of its data strategy are two: first, using large-scale sales to place embodiments in universities and research institutions worldwide, leveraging external developers to form a data-and-skills ecosystem; second, accessing NVIDIA's data and simulation pipeline through the GR00T collaboration. Quantitative metrics of its autonomous data loop (daily collection volume, dataset scale) are missing from public information; marked [To be filled].

4.3. Product Line and Shipment Structure Diagram

图 1-1｜宇树科技产品线矩阵与关键节点时间线

示意与数据来源：招股书披露口径、新浪财经 WRC 2026 观察，信息截止 2026-09-12。

5. Harness Design

Unitree's public engineering system is examined layer by layer against the Harness six-layer model in the parameter card. Overall judgment: Unitree's strengths lie in L2 (embodiment execution) and L3 (fleet orchestration), while its weaknesses lie in L4 (memory) and L6 (governance) — the latter has virtually no public information.

5.1. L1 Context Engineering Layer

WVLA2.0 and G1's full-chain voice interaction (natural language → semantic understanding → spatial perception → trajectory planning → action execution) are the main evidence for L1: multimodal inputs are organized into task context executable by the model. In the H2+ reference design, the GR00T platform provides perception–language–action context assembly capability. However, Unitree has not disclosed engineering components of context engineering such as retrieval, compression, and caching; the rating is mechanism present, granularity not visible.

5.2. L2 Tools and Execution Layer

This is Unitree's strongest layer. Core components (motors, reducers, dexterous hands, 4D LiDAR L2) are self-developed and self-produced; the embodiment's execution precision and dynamic range (360 N·m joint torque, side flips, catapult flips over 3 m) are backed by a large number of public demonstrations. The "App Store" announced in 2025-12 means the execution layer is beginning to be platformized: third-party skills are registered into the embodiment in application form, similar to a consumer-grade version of the L2 tool registration mechanism.

5.3. L3 Orchestration and Control Layer

The three CCTV Spring Festival Galas (2021 A1 "Niulai" (Rise Up), 2025 H1 "Yangge BOT", 2026 G1/H2 "Wushu BOT") are the largest-scale, highest-exposure public validation of multi-robot orchestration among all vendors in this group: 16 H1 units performed yangge in sync with real performers, and the G1/H2 fleet, with positioning, completed the long staff, nunchaku, drunken fist, and continuous freestyle table-flipping parkour — all fully autonomous. Multi-robot coordinated scheduling, trajectory timing alignment, and exception fault tolerance all require support from L3-level orchestration systems, but Unitree has not made public the technical form of its orchestration framework.

5.4. L4 Memory and State Layer

No disclosure of session memory, long-term memory, or checkpoint mechanisms appears in public information. WVLA2.0's meeting-room tidying task is a single-scenario, short-horizon task and does not yet constitute L4 evidence. The rating is gap, marked [To be filled].

5.5. Evaluation and Observation Layer

The four golds at the Games and the stability of Guinness-level performances are "on-site evaluation" evidence for the motion control layer, but Unitree has not disclosed task success rates, regression test sets, or simulation evaluation pipelines (contrast: Galbot's synthetic data pipeline, AgiBot's GE-Sim closed-loop training). Indirect evidence is the real operating-condition feedback obtained through large-scale mass-production shipments. The rating is indirect evidence, no public system.

5.6. L6 Governance and Safety Layer

No disclosure of permission tiering, auditing, guardrail models, or safety certification appears in public information. As a listed company, its governance obligations cover at least information disclosure and product quality compliance, but public details on robot runtime safety governance (emergency-stop strategies, privilege-escalation protection, data compliance) are missing. This is a common gap for most vendors in this group, and it is particularly pronounced for Unitree. Marked [To be filled].

5.7. Six-Layer Strength/Weakness Summary

6. Case Studies

6.1. Fully Autonomous Fleet Performance at the Spring Festival Gala

Three appearances at the CCTV Spring Festival Gala have been the best amplifier of Unitree's brand and technology: 24 A1 units in "Niulai" (Rise Up) in 2021; 16 H1 units in "Yangge BOT" in 2025 (directed by Zhang Yimou; the official account calls it the world's first large-scale, fully AI-driven, fully autonomous fleet humanoid performance); and "Wushu BOT" on New Year's Eve of the Year of the Horse 2026, with G1/H2 performing synchronized martial arts with real performers from the Tagou Martial Arts School, including the long staff, nunchaku, drunken fist, backflips, continuous freestyle table-flipping parkour, and catapult flips over 3 m. The significance of this case is that it proves fully autonomous multi-robot orchestration can operate stably in a zero-fault-tolerance live broadcast environment, directly corresponding to L3/L5 capability.

6.2. World Humanoid Robot Games

At the 1st World Humanoid Robot Games in 2025-02, the H1/G1 platforms won four gold medals in the 1,500 m, 400 m, 100 m hurdles, and 4x100 m relay. The competition environment provided Unitree with comparable results competed on the same field as rivals, and also explains why the research-and-education market treats Unitree embodiments as the default experimental platform.

6.3. H2+ and the NVIDIA GR00T Reference Design

On 2026-06-01, Jensen Huang announced the H2+ reference design: H2 embodiment + Sharpa five-finger dexterous hand (25 DOF per hand) + Jetson Thor compute + Isaac GR00T platform, about 1.8 m / 68 kg. This collaboration advanced Unitree from "a hardware factory selling embodiments" to "a reference carrier for global humanoid AI development" — a landmark event in the combination of hardware scale and technical ecosystem.

7. Summary

7.1. Strengths and Limitations

Strengths: shipment volume ranked first globally (over 5,500 humanoid units in 2025) with five consecutive years of profitability — the only vendor in this group possessing both scale and profitability; a price band covering 29,900–170,000 yuan, with the shipment structure shifting from research-and-education-led to humanoid-revenue-led; a cost barrier from self-developed core components; and, after the STAR Market listing, dual ammunition of capital and production capacity (IPO proceeds targeting 190,000 units per year).

Limitations: the foundation model layer (WVLA2.0) lacks paper-level technical evidence, and quantitative metrics of the self-built data loop are missing; industrial-grade operation cases (contrast: UBTech's automotive-factory training, Galbot's CATL) are thinly covered in public information; the L4/L6 layers are nearly blank.

7.2. Applicability Boundaries

Unitree embodiments are best suited to research-and-education, cultural-entertainment and performance, competition, developer-ecosystem, and brand-marketing scenarios. If normalized operation delivery in industrial scenarios is required (automotive-factory training, production-line logistics), prioritize vendors that already have named customer cases, such as UBTech, Galbot, and AgiBot; if a "brain-as-model" procurement is needed, Unitree is not a model-layer vendor (contrast: PI, see 08-physical-intelligence.md for details).

7.3. Selection Recommendations

1. Universities and research institutions should first consider the G1/R1 price tier, paired with GR00T or a self-developed control stack.
2. Performance and cultural-entertainment projects can directly procure H1/G1/H2, whose fleet orchestration capability has been validated by the Spring Festival Gala.
3. Watch the H2+ reference design route: if a team develops on the NVIDIA GR00T stack, H2+ is one of the carriers with the least engineering friction today.
4. Before industrial procurement, the vendor must be required to provide operating-condition data and a safety governance plan — public information is insufficient on this dimension for Unitree and must be supplemented in the commercial process.

Information Gap Statement

1. Specific H2 pricing: no official figure was found; marked [To be filled].
2. Two-metric conflict: 2025 revenue of 1.699 billion yuan vs. 1.708 billion yuan (both different reprinted metrics of the prospectus), and net profit excluding non-recurring items of 591 million yuan vs. 600 million yuan — this document presents them as ranges side by side without choosing one.
3. L4/L6 layers (memory mechanisms, runtime safety governance) are missing from public information; the ratings are based on the absence of public information rather than the absence of capability.
4. WVLA2.0 lacks paper-level technical sources; architecture descriptions follow the launch events and media reports.
5. Overseas after-sales system: few public details; the delivery and O&M model behind the overseas revenue share of over 40% is undisclosed.

8. References

1. Unitree official website (product matrix: H2/A2/R1/G1/Go2/B2/As2/L2) — Unitree, 2026. https://www.unitree.com/?id=1
2. Unitree News Center (Spring Festival Gala / Cyber Kung Fu / H2 Plus GR00T announcements) — Unitree, 2026. https://www.unitree.com/?p=766/
3. Unitree's STAR Market IPO passed review; H2+ officially announced with NVIDIA — Gasgoo / Tencent News, 2026. <https://gu.qq.com/resources/shy/news/detail-v2/index.html?t=1#/index?_tentrees_trans=0&id=SN2026060122370194ff88c2>
4. IPO proceeds targeting annual capacity of 190,000 units (75,000 humanoids + 115,000 quadrupeds) — Tencent News via The Paper, 2026. <https://gu.qq.com/resources/shy/news/detail-v2/index.html?t=1#/index?_tentrees_trans=0&id=SN2026060122025197646cbe>
5. Unitree's third Spring Festival Gala appearance, "Wushu BOT" — Jingbao Net, 2026-02-16. https://news.bjd.com.cn/2026/02/16/11588316.shtml
6. Unitree humanoid robot (history / financials / IPO data) — Baidu Baike, 2026. https://baike.baidu.com/item/%E5%AE%87%E6%A0%91%E4%BA%BA%E5%BD%A2%E6%9C%BA%E5%99%A8%E4%BA%BA/67401754
7. Unitree H2 entry — Chinese-language encyclopedia (newton), 2026. https://www.newton.com.tw/wiki/Unitree%20H2/67392987
8. WRC 2026 observation: the industry enters the industrialization-landing phase (Unitree listing / shipments / price band) — Sina Finance / Global Robotics Information, 2026-09-12. https://www.toutiao.com/article/7684438007058842150/
9. AgiBot: leading full-stack embodied-AI vendor (in-depth report, including Omdia's 2025 global shipment baseline) — Zheshang Securities, 2026-03-28. https://www.sgpjbg.com/info/df7be3e7242d311ece3785906662f44c.html