CITIC Securities: Embodied intelligence opens up growth space for humanoid robots

According to the Zhitong Finance APP, Guotai Junan has released a research report stating that the rapid development of humanoid robots is driven by embodied intelligence, which is the core factor for commercialization. The improvement in robot intelligence levels and real-time control performance will drive the growth in demand for perception performance, computing power, and communication efficiency, benefiting edge-side sensing, control, and communication chips. The realization of embodied intelligence opens up growth space for humanoid robots, with broad future application prospects, driving the performance of complete machine manufacturers upward. The release of high-performance multimodal large models and breakthroughs in motion algorithm performance will act as catalysts. Recommended targets: Fengcai Technology (688279.SH), Orbbec-UW (688322.SH), Triductor (688259.SH), HYC (688001.SH).

Guotai Junan's main viewpoints are as follows:

Market space exceeds one trillion, achieving embodied intelligence is the core of commercialization

Humanoid robots possess perception methods, body structures, and movement styles similar to humans, making them highly adaptable to human society. Potential application scenarios can cover production manufacturing, social services, hazardous operations, and more. According to the "Humanoid Robot Industry Development Research Report (2024)," from 2024 to 2028, the overall intelligence level of humanoid robots in China will still be at Lv1, with a few of the latest products exploring Lv2, and the market scale will be less than 10 billion yuan. As the intelligence level evolves towards embodied intelligence, humanoid robots are expected to break through specific scene and task limitations, achieving comprehensive coverage across various industries, with the market scale of humanoid robots in China expected to exceed one trillion yuan after 2045.

Multimodal and reinforcement learning enhance operational control performance, continuous innovation in perception sensing, and upgrades in communication and computing power enhance real-time control

Multimodal large models are key to improving human-machine interaction efficiency and enhancing situational understanding. Currently, NVIDIA's GR00T and Tesla's Grok3 large models are actively integrating multimodal perception to accelerate the improvement of robot interaction and decision-making accuracy.

Chain of Thought empowers large models for efficient reasoning. GPT-4.5 has passed the Turing test to a certain extent. Currently, UBTECH is developing an embodied reasoning large model based on DeepSeek-R1, which is expected to achieve accurate and efficient responses and decisions in complex environments.

Motion control algorithms are core to coordinating the execution of full-body movements. Reinforcement learning can achieve efficient learning of movements such as gait and running based on reward functions, enhancing generalization capabilities, and is expected to become the main paradigm for motion algorithms.

Pure visual solutions significantly reduce hardware costs and achieve efficient visual perception. Six-dimensional force sensors and electronic skin can effectively improve sensitivity, helping robots accurately understand environmental information and execute actions precisely. The combination of pure vision, six-dimensional force sensors, and electronic skin is expected to become the standard sensor solution in the future. Real-time control requires efficient communication protocols and powerful hardware computing power as underlying support. EtherCAT features high real-time performance, low latency, and high synchronization, reducing communication latency from the millisecond level of the CAN protocol to the microsecond level, and is expected to become the mainstream communication protocol for robots. As robot intelligence evolves towards embodied intelligence, the demand for edge-side computing power is expected to continue to grow, driving performance upgrades in edge-side chips Risk Warning

The speed of technological iteration is slower than expected, and the progress of localization in core areas is not as anticipated