Here Come Humanoid Robots: Industry Makes Clear Pivot Toward “Dedicated-Purpose” Commercial Deployments

Goldman analyst Jacqueline Du published her field trip notes after speaking with six C-level management teams at robotics companies across Hangzhou, Shanghai, and Shenzhen in recent days. The companies she met with include Unitree, Mechmind, Fourier, LimX Dynamics, UBTech, EngineAI, Paxini, and Orbbec.

“Overall, we see the humanoid robot industry is experiencing a clear pivot towards ‘dedicated-purpose’ commercial deployments,” Du told clients on Thursday morning.

Home humanoid robots are getting closer.

Shenzhen MindOne Robotics is testing their robot brain on the Unitree G1, and it looks like the G1 has already learned to do human-like household chores.
Watering plants, moving packages, cleaning mattresses, tidying up, etc.

Honestly,… pic.twitter.com/hPddAaQ0hd

— CyberRobo (@CyberRobooo) November 14, 2025

She continued: “This strategic focus leverages current, achievable task planning, mobility and interaction capabilities, leading to more reliable and immediate deployment in specific vertical applications. Examples include security patrolling, guiding/guest services in public venues, and logistics tasks such as pick-and-place and sorting simple items in factories.”

Du said humanoid robot shipments are set to accelerate from estimated 2025 global volumes of around 15,000 to 20,000 units to “multi-fold increases in shipment volumes into 2026-2027E.”

Here are more of Du’s key takeaways from the AI robotics field trip in China:

Targeting a multi-fold increase in shipment volumes into 2026-2027E

Based on the collective feedback from the major humanoid robot players and key supply chain companies we met on this trip, we believe global humanoid robot shipments in 2025 might have reached around 15,000-20,000 units vs. GSe of 20,000 (report link) and third-party data of 13,000-16,000 (see link and link), with Chinese players contributing to the bulk of these volumes at the moment, which came from scientific study, robotics AI training, education, entertainment/stage performance, and data factory demand last year. At such an early and nascent stage of the humanoid robot market, the exact shipment figures are less critical than the overall growth trajectory and the rapid pace of technological development in our view. For the 2026-2027 outlook, these leading humanoid robot manufacturers anticipate multi-fold increases in shipments. Following the 2025 figures, which range from several hundreds to thousands of units, their targets for 2026 and 2027 are set in the thousands to tens of thousands. This projected acceleration is underpinned by an increasingly mature supply chain, optimized cost curves, and expanding application opportunities. However, achieving these targets is anticipated to be challenged by the imperative for robust production consistency and the implementation of novel, multi-stage testing protocols inherent to this nascent industry.

Encouraging progress in motion control with rapid iterations

We watched the live demos of the published products of these humanoid robot makers, which in our view revealed significant advancements in motion control, exhibiting enhanced robustness and flexibility across both wheeled upper-body platforms and full bipedal systems. This represents a substantial improvement compared to their performance in the preceding year. One manufacturer asserted the achievement of ‘cerebellum-level’ whole-body control, substantiated by two practical evaluation criteria: the robot’s ability to navigate previously unmapped terrains and its capacity for comprehensive remote control of the entire body, rather than segmented upper or lower body control. Furthermore, insights from multiple companies indicate an accelerated product iteration cycle for humanoid platforms, now reaching approximately 6-8 months per generation. This rapid iteration is largely attributed to an 80-90% in-house component design capability, which is crucial for ensuring seamless hardware-software integration and optimizing their respective performance’ upper limits’ within condensed R&D and testing cycles.

Clear pivot towards “dedicated-purpose” commercial deployments

The current reliance on simulated and synthetic data for pre-training with a significant ‘sim-to-real’ gap remains a challenge, with simulated accuracies of 80-90% often collapsing to below 50% in real-world scenarios. As both large-scale high quality real-world data collection and world model approach requires time, leading Chinese humanoid robot developers are increasingly prioritizing “dedicated-purpose” commercial deployments. These applications, such as security patrolling and guest services in public venues (e.g., hotels, banks, museums, exhibition centers, auto dealerships, and supermarkets), effectively utilize existing task planning, mobility and interaction capabilities while circumventing the complexities of highly dexterous manipulation. Within industrial applications, the utility of humanoid robots requiring dexterous hands or grippers is presently confined to logistics tasks like box movement and simple item sorting, primarily due to AI limitations in addressing unpredictable corner cases in factory environments.

Advancing robot intelligence through hybrid AI and data strategies

In the near term, humanoid robot manufacturers are increasingly standardizing their approach by integrating with established Large Language Model (LLM) and Vision-Language Model (VLM) stacks, such as those offered by Alibaba (Qwen), Doubao, and Tencent. This strategic alignment positions proprietary data engines as the critical differentiator for developing deployable robotic intelligence. High-quality real-world data is identified as the primary constraint bridging the gap between mature hardware technologies and scalable, practical applications. Consequently, companies are engaged in a ‘data recipe’ arms race, with differentiation driven by their target end-applications. While all robot OEMs are pursuing data collection strategies, they converge on varying mixes of three primary data inputs: (1) teleoperated human or expert demonstrations, which offer high control but are typically expensive for imitation learning; (2) simulation, which is cost-effective per additional sample but suffers from imperfect realism; and (3) real-world video datasets, which provide the highest data availability but may exhibit poor translation to real-world accuracy. From this trip, increasingly we heard more mentions of the world model approach which may potentially empower robots with a form of common sense about their environment, moving them beyond reactive behaviors to proactive, intelligent agents capable of complex planning and adaptation.

Differentiated profit models for 2C and 2B markets

A range of profit models has emerged, broadly categorized by their target markets: 2C (business-to-consumer) and 2B (business-to-business) applications.

For companies targeting 2C applications, the primary focus is on delivering differentiated functionality and enhanced user experience. This often involves emphasizing “emotional value” and capturing specialized vertical niches where unique features or interactions can command a premium. The goal is to create a product that stands out through its capabilities and the user’s engagement with it.

In contrast, companies targeting 2B applications anchor their pricing strategies to the customer’s Return on Investment (ROI). This typically involves demonstrating how the robot can improve throughput, enhance efficiency, or reduce labor costs. For instance, UBTech has indicated that in sorting and logistics applications, customers are willing to invest in robots once they achieve approximately 50% of a human worker’s throughput. This level of performance can lead to a payback period of around two years, assuming a run-time of about 10 hours per day. Even a three-year payback period is considered acceptable by customers operating in particularly labor-constrained environments, highlighting the value proposition of automation in addressing critical operational challenges.

Investment implications: We recommend being selective; Buy Sanhua H and Sell Moon’s Electric

2026 overall may shape up to be a critical “proof-of-volume and expectation-reset” year, as we believe investors likely will continue to value key supply chain stocks on (i) whether milestone volume expectations (e.g., the “one million robots” narrative) get revised up or not, which is likely driven by the evolution pace of AI generalization capability or effective data/model strategies; and (ii) evolving market share and content per robot for individual supply chain companies. Given high market optimisms and long run expectations have been baked into the current share price per our tests (see Sanhua Intelligent Controls (002050.SZ): Downgrade Sanhua A to Neutral on recent outperformance; expectations for humanoid robots are too high, too soon and Moons’ Electric (603728.SS): D/G to Sell on continuously evolving dexterous hand technology roadmap and narrowed opportunity for coreless motor), we recommend staying selective. Among our coverage, we are Buy-rated on Sanhua H, Inovance and Shuanghuan, Neutral-rated on Sanhua A, LeaderDrive and Best Precision. We are Sell-rated on Moon’s Electric.

Separately, in the U.S., Tesla said it is on track to begin volume production of the Cybercab in 2026, with Optimus humanoid robot output “hopefully” starting toward the end of the year.

Next week, we are expected to conduct our own field trip of AI robotics and counter-AUS companies at undisclosed locations. We’ll see how that turns out – watching the rise of Skynet in real-time.