UBTech Market Research

1. Introduction

UBTech's within-group differentiation main line is "automaker training + multi-scenario delivery": as the HKEX "first humanoid robot stock" (9880, listed before Unitree's A-share listing), it is the vendor among this group's ten with the longest list of industrial customers and the widest scenario coverage — BYD, NIO, Zeekr, FAW-Volkswagen, Audi FAW, Dongfeng Liuzhou Motor, BAIC New Energy, Foxconn, SF Express, Sany Renewable Energy, then Airbus, Texas Instruments, and the China-Vietnam border port. In FY2025 it delivered 1079 full-size humanoids (audited basis), making it the only vendor in this group with audited delivery data.

1.1. Vendor Overview

1.2. Development History and Financial Basis

1.3. Business Model and Pricing

UBTech's revenue structure is complete-machine sales + scenario solutions + data collection services, with customers covering five categories: automakers, 3C electronics, aviation manufacturing, semiconductors, and smart logistics. The Walker S2 enterprise quote basis is about $160,000 (not an officially published price); the order basis is "Walker S orders and reported factory training orders exceed 800 million yuan" (RobotTesters basis). There is a two-basis conflict in capacity planning, presented side by side in this document: Basis 1 (Science and Technology Daily / UBTech official basis) — 2026 humanoid robot capacity reaches 10,000 units, with 1-million-unit preparation from 2028; Basis 2 (Robozaps) — 2026 capacity guidance of 5000 units, ramping to 10,000 units in 2027.

2. Glossary

3. Feature Description and Product Line

3.1. Product Line Overview

The core differentiation of the Walker S2 is runtime continuity: dual-battery hot-swap, with the robot autonomously deciding "swap or charge", completing the swap in 3 minutes — turning "endurance" from a robot parameter problem into a scheduling problem, approaching the shift-change logic of human workers.

3.2. BrainNet 2.0 and Co-Agent

BrainNet 2.0 is the foundation of UBTech's robot software stack; the Co-Agent on top of it (officially called the world's first collaborative agent for industrial humanoids) handles multimodal reasoning, embodied interaction, and skill model invocation, forming a "task understanding → planning → execution → exception handling" closed loop, and supports multi-robot edge-side coordination. This is the software layer closest to the "agent runtime" concept among this group's Chinese vendors.

4. Platform Architecture

4.1. Software-Hardware Stack

UBTech's stack has four layers: the body layer (52-degree-of-freedom Walker S2 + fourth-generation dexterous hand + servo drive self-research tradition); the control layer (full-body control + energy management for 3-minute battery swap); the software layer (BrainNet 2.0 multi-robot coordination); the agent layer (Co-Agent task closed loop). The patent reserve (2790 patents, world's number one in valid humanoid patents) is mainly accumulated in the body and servo layers.

4.2. Factory Training and Data Collection System

UBTech's landing methodology is a two-year "POC → normalized" path: first enter automakers with small-scale training (exchanging operating conditions for free or low cost), then transition to paid scaled delivery after accumulating operational data and scenario understanding. The Zigong data collection center (159 million yuan contract) makes data collection itself an external service — the "data industry" dimension of industrial humanoids.

4.3. Five-Scenario Deployment Diagram

图 9-1｜优必选五大场景与代表性部署（信息截止 2026-09-12）

数据来源：科技日报、优必选公开披露与第三方档案，信息截止 2026-09-12。

5. Harness Design

5.1. L1 Context Engineering Layer

Co-Agent's multimodal reasoning handles the context assembly for industrial tasks: production-line station information, material status, and work-order instructions are organized into an executable task context. The context for industrial scenarios is mainly structured work orders, and is inherently more controllable than open scenarios.

5.2. L2 Tooling and Execution Layer

52 degrees of freedom + fourth-generation industrial-grade dexterous hand + 15kg payload constitute the execution layer; millimeter-level door-lock QC, 99.7% sorting accuracy, and autonomous bolt-sleeve removal provide multi-dimensional evidence of execution precision and complexity. The 3-minute battery swap is an architectural-level design for execution continuity — this group's most engineered solution to "runtime continuity".

5.3. L3 Orchestration and Control Layer

BrainNet 2.0 + Co-Agent constitute an explicit orchestration layer: a task understanding-planning-execution-exception handling closed loop + multi-robot edge-side coordination (multi-robot scheduling that does not depend on the cloud). This is the only publicly named multi-robot coordination software product among this group's Chinese vendors.

5.4. L4 Memory and State Layer

The Zigong data collection center (159 million yuan contract) accumulates training data into a manageable asset; the operating-condition data accumulated over two years of automaker training is a task-level "experience library". The runtime memory mechanism is not publicly disclosed and is marked [To be filled].

5.5. Evaluation and Observability Layer

UBTech's evaluation evidence is audit-grade delivery + customer-level KPIs: 1079 units is on the audited basis (unique in this group); BYD's doubling of sorting efficiency, 99.7% accuracy, non-destructive QC >99%, etc., are quantified results for customer scenarios. However, there is a lack of published standardized performance benchmarks (such as same-task comparisons with competitors), and RobotTesters' commentary points out that it has no independent performance verification — under a closed-source software stack, it is difficult for third parties to reproduce the tests.

5.6. L6 Governance and Safety Layer

As an HKEX-listed company, its L6 foundation is listing compliance and the audit system; the 2790 patents and leading / deep participation in multiple international and national standards (Science and Technology Daily basis) are governance assets at the level of industry-standard discourse power; the first batch of customs and border inspection deployments implies a national-level safety review endorsement. The public details on robot runtime safety (emergency stop, human-robot collaboration safety standards) are still limited; the U1 consumer series controversy (2026-07 weekly report basis) highlights the quality control and public opinion risks in its product line expansion.

5.7. Six-Layer Strengths Summary

6. Actual Cases

6.1. Automaker Training: From POC to Normalization

UBTech's automaker training has gone through a two-year POC → normalization path, with named customers and results: NIO (door-lock QC, millimeter-level precision), Zeekr (5G smart factory automated warehouse tote transport), BYD (Walker S1 sorting accuracy 99.7%, sorting efficiency doubled, non-destructive QC accuracy >99%), Sany Renewable Energy (5G wind power smart factory: autonomous bolt-sleeve removal, sorting and stacking, full-pallet transport); orders cover Geely / Zeekr, FAW-Volkswagen (Qingdao), Audi FAW, Dongfeng Liuzhou Motor, BAIC New Energy, Foxconn, etc., with a single order of 250 million yuan setting a record in 2025-09. The 1000th Walker S2 rolled off the Dongfeng Liuzhou Motor production line (end of 2025-12), marking the node of training transitioning to mass production. The specific deployment unit counts for each automaker are all undisclosed and are marked [To be filled].

6.2. Airbus and Texas Instruments

In 2026-01, Walker S2 procurement entered the Airbus aircraft manufacturing scenario: parts logistics, tool management, inspection support (early POC, unit count and amount undisclosed). The Texas Instruments (TI) semiconductor line deployment includes the Sherman SM1 fab, and the contract includes a reverse component procurement clause. These two cases expand UBTech's scenario boundary from automotive to aviation manufacturing and semiconductors — the two categories with the highest requirements among the five scenarios, and also prove the portability of the Walker S2 platform.

6.3. Border Port Deployment

The China-Vietnam border port project in Fangchenggang, Guangxi (about 264 million yuan contract): by mid-2026 it was already operating in the real port environment, handling traveler guidance, crowd management, patrol, and cargo inspection; it is the first batch of deployments of humanoid robots in customs and border inspection scenarios, and reportedly there are plans to extend to airports / ports / seaports. The value of this case lies in institutional-level verification in a non-factory scenario — port deployment requires passing the safety and process reviews of customs and border inspection agencies, which is an on-the-ground endorsement for L6 governance.

7. Summary

7.1. Strengths and Limitations

Strengths: this group's longest list of named industrial customers and widest scenario coverage (five major scenarios + port); the only audited-basis delivery data (1079 units); the continuity problem solved by the 3-minute battery swap is a core pain point of industrial procurement; the intellectual property barrier built by the BrainNet 2.0 / Co-Agent software stack and 2790 patents; the financing and governance transparency of the HKEX-listed platform.

Limitations: closed-source software stack + no independent performance verification (absence of third-party evaluation); capacity basis conflict (10,000 units vs 5000-unit ramp); the deployment unit count for each customer is opaque; the U1 consumer line controversy; FY2025 still has a loss of 790 million yuan, and the profitability model for the humanoid business is not closed.

7.2. Applicable Boundaries

Suitable for scaled procurement in automotive, 3C electronics, semiconductors, aviation manufacturing, smart logistics, and port / patrol-type scenarios — especially when the buyer needs multi-scenario platform reuse (the same Walker S2 platform across scenarios) and systematic delivery capability. For scientific research and education, and cultural entertainment and performing arts scenarios, Unitree should be chosen; for home caregiving scenarios, Fourier should be chosen; teams with high requirements for openness (open source, capable of self-developed takeover) need to note the constraints of UBTech's closed-source stack.

7.3. Selection Recommendations

1. Contract with "scenario KPI commitment + audit-grade acceptance": UBTech's customer KPI data (99.7% sorting accuracy, etc.) should be converted into contract metrics rather than promotional material.
2. Require third-party reproduction testing or open evaluation interfaces — in response to the industry commentary of "no independent performance verification", the buyer should build its own acceptance tests.
3. Schedule procurement for capacity and delivery based on the conservative value of the two bases (2026 5000-unit ramp), treating the 10,000-unit basis as the optimistic scenario.
4. Multi-scenario group customers can reference its "same platform across five scenarios" deployment model to reduce operations and maintenance complexity, but must conduct process adaptation evaluation separately for each scenario.

Information Gap Statement

1. Two-basis capacity conflict: 2026 10,000-unit capacity (official / Science and Technology Daily basis) vs 5000-unit ramp (Robozaps third-party basis), presented side by side.
2. Deployment unit counts for each automaker (BYD / Zeekr / NIO, etc.) are all undisclosed and are marked [To be filled].
3. Specific facts of the U1 consumer series controversy have not been expanded through retrieval (2026-07 weekly report basis); this document only provides a risk warning.
4. Walker S2 official pricing has not been published (about $160,000 is the enterprise quote basis); the original URL of the FY2025 HKEX announcement has not been obtained.
5. Airbus / TI deployment unit counts and amounts are undisclosed; the runtime memory mechanism (L4) has no public details.

8. References

1. UBTech: Deeply Cultivating Core Technology, Honing Exclusive Skills — Science and Technology Daily (Toutiao reprint), 2026-03. https://www.toutiao.com/article/7613949193794716160/
2. UBTech Walker S2 — AI Wiki (customers / Airbus / Texas Instruments / port), 2026. https://aiwiki.ai/wiki/ubtech_walker_s2
3. UBTech Walker S2: The Humanoid That Swaps Its Own Battery — RobotTesters, 2026. https://robottesters.com/article-ubtech-walker-s2.html
4. UBTech Walker S2: Price, Verified Specs & 3-Minute Battery Swap (2026) — Robozaps, 2026-07. http://blog.robozaps.com/b/ubtech-walker-s-review
5. Robots Working in Factories (UBTech chapter) — Solar System Science Popularization (adapted from Science and Technology Daily), 2026. https://solarnews.cn/News/Info/15490
6. AI Robot Weekly (UBTech controversy / Unitree IPO / Zhiyuan 15,000 units) — NetEase News, 2026-07. https://c.m.163.com/news/a/L1VPHI9505118I96.html
7. UBTech FY2025 Annual Results Announcement — UBTech / HKEX Disclosure Easy, 2026 (the original announcement URL has not been obtained; listed by "document name + institution + year").
8. Zhiyuan: Full-Stack Leader in Embodied Intelligence (Omdia 2025 global shipment base data chapter) — Zheshang Securities, 2026-03-28. https://www.sgpjbg.com/info/df7be3e7242d311ece3785906662f44c.html
9. WRC 2026 Observation: The Industry Enters the Industrialization Landing Period — Sina Finance / Robot Global Information, 2026-09-12. https://www.toutiao.com/article/7684438007058842150/