Goldman Sachs China Humanoid Robot Research: The industry shifts from "general imagination" to "specific implementation," with 2026 likely to welcome "volume validation + expectation reset"

Goldman Sachs' latest research shows that the humanoid robot industry in China is undergoing a strategic transformation from "general imagination" to "specific implementation." This pragmatic approach, combined with significant advancements in motion control capabilities and rapid iteration cycles, is driving major manufacturers to set shipment targets for 2026-2027 at several times the volume of 2025. According to the Wind Trading Desk, Goldman Sachs analyst Jacqueline Du pointed out in a recent report that during a survey of eight humanoid robot and industry chain companies, including Yushu Technology, UBTECH, Fourier, and Yundongchu, conducted from January 15 to 20, the team observed that the industry is shifting its focus from pursuing general capabilities to vertical scenarios that utilize existing task planning, mobility, and interaction capabilities, such as security patrols, public place guidance services, and factory logistics sorting. Based on feedback from major manufacturers and supply chain companies, Goldman Sachs expects global humanoid robot shipments to reach approximately 15,000 to 20,000 units in 2025, with Chinese companies contributing the majority of the shipments. Looking ahead to 2026-2027, leading manufacturers expect to achieve several times growth, increasing from hundreds to thousands of units in 2025 to thousands to tens of thousands of units. Goldman Sachs noted that 2026 could become a key year for "volume validation + expectation reset," where investors will focus on whether milestone expectations such as "one million robots" are adjusted, as well as the market share and unit value evolution of individual supply chain companies. Aggressive Shipment Targets, Capacity and Testing Challenges According to Goldman Sachs' research, global humanoid robot shipments are expected to be between 15,000 and 20,000 units in 2025, close to Goldman Sachs' previous expectation of 20,000 units, and consistent with third-party data's range of 13,000 to 16,000 units. Chinese companies currently contribute the vast majority of shipments, with demand primarily coming from research, robot AI training, education, entertainment performances, and data factories. In this still early-stage market, leading manufacturers have set ambitious growth targets for 2026-2027. Based on 2025 shipments ranging from hundreds to thousands of units, companies have set their targets for 2026-2027 at thousands to tens of thousands of units, indicating several times growth. Factors supporting this growth expectation include an increasingly mature supply chain, optimized cost curves, and expanded application scenarios. However, Goldman Sachs pointed out that achieving these targets will face challenges such as ensuring production consistency and the multi-stage testing processes inherent in this emerging industry. Significant Improvements in Motion Control, Iteration Cycle Reduced to 6-8 Months During on-site product demonstrations, Goldman Sachs analysts observed substantial progress in humanoid robots' motion control capabilities, whether in wheeled upper-body platforms or fully bipedal systems, showing greater robustness and flexibility, with significant improvements compared to the previous year. One manufacturer claimed to have achieved "cerebellum-level" full-body control capabilities and provided two practical evaluation criteria: the robot can navigate uncharted terrain and can achieve full-body remote control rather than segmented control of upper and lower body. Supply chain integration capabilities are accelerating product iteration. Multiple companies revealed that the product iteration cycle for humanoid robot platforms has been shortened to about 6-8 months per generation. This rapid iteration is largely attributed to the 80%-90% self-designed components, which are crucial for ensuring seamless integration of hardware and software and optimizing performance limits within compressed R&D and testing cycles Application Focus on "Dedicated Landing" to Bypass the Challenges of Dexterous Operations The gap between "simulation and reality" remains a bottleneck in the industry. Currently, robot pre-training heavily relies on simulated and synthetic data, with accuracy rates of 80%-90% in simulated environments often dropping below 50% in real-world scenarios. Due to the time required for large-scale collection of high-quality real data and world model methods, leading humanoid robot developers in China are prioritizing the development of "dedicated" commercial deployments. These application scenarios include security patrols and guidance services in public places such as hotels, banks, museums, exhibition centers, car dealerships, and supermarkets, effectively utilizing existing task planning, mobility, and interaction capabilities while avoiding the complexities of highly dexterous operations. In industrial applications, humanoid robots that require dexterous hands or grippers are currently limited to logistics tasks such as box moving and simple item sorting. This is mainly due to the limitations of AI in handling unpredictable edge cases in factory environments. According to UBTECH, customers are willing to invest in sorting and logistics applications when robots reach about 50% of human worker productivity, which can lead to a payback period of approximately two years (assuming about 10 hours of operation per day). Even in environments with particularly tight labor, a three-year payback period is considered acceptable. Data Strategy Becomes Core Competitiveness, World Models Gain Attention Recently, humanoid robot manufacturers have increasingly adopted standardized approaches, integrating with mature large language models (LLM) and vision-language model (VLM) technology stacks from companies like Alibaba (Tongyi Qianwen), Doubao, and Tencent. This strategy makes proprietary data engines a key differentiating factor in developing deployable robotic intelligence. High-quality real-world data is seen as the main constraint in bridging the gap between mature hardware technology and scalable practical applications. Therefore, companies are engaged in an arms race for "data recipes," differentiated by their target end applications. While all robot manufacturers are pursuing data collection strategies, they adopt different combinations of three main data inputs: remote-operated human or expert demonstration data, which is highly controlled but often expensive; simulated data, which has low costs per additional sample but imperfect authenticity; and real-world video datasets, which have the highest data availability but may convert to real-world accuracy poorly. Goldman Sachs found in this research that the mention of world model methods is increasing, which may endow robots with some common sense about their environment, allowing them to shift from reactive behaviors to proactive agents capable of complex planning and adaptation. Differentiated Business Models: 2C Focuses on Experience, 2B Looks at ROI Different target markets have spawned differentiated profit models, primarily divided into 2C (consumer-facing) and 2B (business-facing) applications. Companies targeting 2C applications mainly focus on providing differentiated features and enhancing user experience, often emphasizing "emotional value" and capturing specialized vertical niches, achieving premiums through unique functionalities or interactions. The goal is to create products that stand out through capabilities and user engagement. In contrast, companies targeting 2B applications anchor their pricing strategies on customer return on investment (ROI), typically demonstrating how robots can increase productivity, enhance efficiency, or reduce labor costs UBTECH stated that in sorting and logistics applications, when robots reach about 50% of human workers' productivity, customers are willing to invest, which can lead to a payback period of about two years. Even in environments with particularly tight labor, a three-year payback period is considered acceptable, highlighting the value proposition of automation in addressing critical operational challenges