Goldman Sachs: Large-scale Deployment of Chinese Humanoid Robots Expected Between 2027-2029

Goldman Sachs reports that China's humanoid robotics industry is accelerating towards commercialization, but patience is required from proof-of-concept to large-scale deployment.

From May 18th to 22nd, Goldman Sachs analysts intensively visited 14 robotics companies in Hong Kong, Shenzhen, and Beijing, covering listed companies and targets in the primary market. Goldman Sachs believes that progress in multimodal AI technology stacks and data acquisition systems indicates that the industry is "a few steps closer" to actual large-scale deployment, but most companies expect true large-scale commercial use to be achieved between 2027 and 2029.

The current industry commercialization is still mainly in the proof-of-concept (POC) stage, with orders in industrial and logistics scenarios mostly for small-batch pilots. Goldman Sachs points out that from POC to batch deployment usually requires 3 to 6 months of verification, under 50 units small-batch testing, about 12 months of assessment cycle, and then it may proceed to a pilot scale of 50 to 100 units per customer.

AI Models: From Single VLA to Multimodal Integration

Goldman Sachs observes that the industry's discussion on models has shifted from the single VLA (visual-language-action) framework to a fusion architecture of VLA and world models. World models are seen as a functional complementary layer to action models, used for state prediction, action verification, and planning enhancement in uncertain environments. Companies such as Galaxy General, Galbot, Lingchu Intelligence, and Wanxun Technology all clearly proposed similar fusion routes.

The scale of model parameters is also on the rise, with some companies starting to explore training solutions with 40 to 80 billion parameters, a significant increase from the previous few billion parameter-based systems. However, many companies frankly admit that models still need several rounds of iteration to achieve stable quality levels suitable for deployment.

Data: High-Quality Real Data Remains the Biggest Bottleneck

High-quality real-world data continues to be the core bottleneck constraining deployment. Industry consensus is shifting from general "data recipe" discussions to building scalable data acquisition architectures. Human-centered collection methods—including wearables and first-person perspective recordings—are becoming mainstream choices.

Pasini currently operates 5 data factories nationwide, Galaxy General aims for a data scale of 1 million hours this year, Lingchu Intelligence has deployed over 800 robots for continuous data return. UBTECH's management expects the demand for data factories to remain strong or further increase in 2026. Several companies guide that the proportion of data-related revenue in 2026 will increase.

Commercialization and Costs: Wheeled Robots First, Cost Reduction Depends on Scale

In terms of landing form, many companies prefer the combination of wheeled chassis with two to three-finger grippers, believing that this solution can cover 70% to 90% of industrial application scenarios, while bipedal humanoid robots are more seen as a long-term direction. Yuejiang Technology reveals that the average price of its humanoid robots is about 300,000 RMB, with a gross margin of 45%; UBTECH's industrial humanoid robots' average price guidance for 2026 is 550,000 to 650,000 RMB, down from 700,000 to 800,000 RMB in 2025.

On the cost side, UBTECH's material costs have dropped from about 400,000 RMB at the beginning of 2025 to just above 200,000 RMB, with a goal of reducing to about 100,000 RMB by around 2027. Goldman Sachs summarized that scale effect and full-stack self-research are still the main paths for cost reduction currently.