Embodied robotics remains an industry in its nascent stages, yet the future outlook is positive with significant progress anticipated at the industrial level, making it a sector worthy of ongoing investor attention. The supply chain is undergoing a sifting process where genuine, high-quality components will retain their investment value post-adjustment. Within the domestic supply chain, priority should be given to high-quality assets preparing for listing, followed by core component suppliers with favorable market structures and high value density. Investors are advised to actively monitor original equipment manufacturers (OEMs), as well as segments with strong competitive landscapes like machine vision and harmonic reducers. The key observations are as follows:
The recent World Artificial Intelligence Conference (WAIC) held in Shanghai saw participation from numerous companies specializing in embodied intelligence hardware, models, and components. A central challenge for embodied AI companies currently is demonstrating credible commercial viability. This year's on-site demonstrations trended towards greater precision, higher generalization capabilities, and longer-horizon operational scenarios. Concurrently, the dominant model architecture has evolved from last year's Vision-Language-Action (VLA) models to a "World Model + VLA" approach. Demand for Ego/UMI data collection equipment and the number of suppliers in this space have increased noticeably.
OEM and Product Landscape Expands
By the first half of 2026, the number of integrated robotic products in China reached approximately 400, an increase of 70 from the end of 2025. While producing prototypes is relatively straightforward, the real challenges lie in hardware—ensuring consistency in mass production and improving operational lifespan—and in software, particularly the development of effective embodied large language models. Furthermore, as technological pathways have not yet converged and standards remain ununified, there is a trend of cross-penetration between model developers and OEMs. Model firms are launching their own hardware products to showcase their models' capabilities, while OEMs are also developing proprietary embodied AI models.
Model Architecture Evolution
While most demonstrations last year featured VLA models, this year the majority showcased a "World Model + VLA" architecture. However, implementations vary among companies. Some integrate the world model directly to learn physical laws and enable pre-action simulation, while others develop the world model and VLA as separate, co-existing models.
Data Collection and Training Dynamics
Firstly, data collection methods are diversifying. Last year's exhibitions were dominated by motion capture suits, whereas this year saw an increase in tactile gloves, UMI grippers, and Ego devices (e.g., dual grippers, head-mounted units, smartphones). Some vendors have already secured orders for several thousand units of real Ego and UMI equipment, with prices trending downward. While creating a UMI/Ego data collection setup is not exceptionally difficult, the greater challenge lies in improving data efficiency and expanding the modalities of data captured.
Secondly, regarding training data structure, most companies now employ a five-layer framework comprising internet data, human behavior data (motion capture and ego data), simulation data, teleoperated real-robot data, and autonomous real-world robot operation data. The proportions of each data type vary, but real-robot data remains indispensable for post-training refinement.
Application Scenarios Deepen and Broaden
The application landscape is marked by a push towards finer precision, higher generalization demands, and more long-horizon tasks, with logistics sorting seen as a likely candidate for early real-world deployment.
In industrial settings, applications have become more varied. Last year's focus was primarily on搬运 and loading/unloading, with some logistics sorting. This year introduced more delicate operations, such as screw-driving by CGS General-Purpose Robots, flexible grasping and precision assembly of wire harness connectors by Mech-Mind, and optical module insertion/extraction, inspection, vacuum sealing, and boxing by Lingchu Intelligence in collaboration with Changfei Optical Fiber.
In commercial service scenarios, multiple companies demonstrated shelf-picking (e.g., for pharmaceuticals, retail; by firms like CGS General-Purpose, Mech-Mind, Ant Alpha), hotel laundry services (by Da Xiao Robot, KEENON, Qionge Intelligence), and beverage preparation. Shelf-picking highlights generalization ability, while hotel laundry tasks demonstrate competency in long-horizon operations.
For home environments, CGS General-Purpose Robots demonstrated breakfast preparation capabilities like baking bread, while Fourier showcased human-robot interaction in home settings integrated with external large language models.
Potential risks include slower-than-expected breakthroughs in key technologies and delays or shortfalls in downstream application development and demand.
Comments