Humanoid robotics has entered a new phase driven by embodied intelligence, according to a research report from Shenwan Hongyuan (SWHY). The report highlights that core components form the hardware foundation, with mass production driving demand for reducers, servos, and sensors. As the intelligent core of robots, embodied large models create software barriers through algorithms and simulation platforms. Training data is a critical constraint for model iteration, with multiple approaches—including real-world teaching demonstrations, first-person perspective collection, and synthetic data from simulations—running in parallel, continually increasing the scarcity of data supply. For the medium to long term, the firm recommends positioning along three main lines: ① manufacturers of key humanoid robot components; ② embodied model and robot developers; and ③ companies related to robot data collection hardware and data services.

The report identifies four key factors in the development of the robotics industry: models, data, hardware, and application scenarios. Models serve as the robot's brain, responsible for reasoning and decision-making. Data is the fuel for model training; without high-quality data, effective training is impossible. The hardware is the robot's physical carrier, defining the upper limit of execution capability, as all intelligent instructions ultimately rely on it. Application scenarios are the focus for eventual commercial deployment and determine the source of demand for robots, as well as their task boundaries. Robots must have a scenario to generate value. Shenwan Hongyuan believes that at the current stage, models and data represent the industry's biggest bottlenecks, an area that has received relatively little attention and research from the market.

Mainstream players are continuously iterating their models, yet the model architecture has not yet converged. Key examples include the progression from PI's π0 to π0.7, and from Google's RT-1 to Gemini Robotics 2, achieving breakthroughs in cross-hardware generalization, operational fluency, and data generation bottlenecks. While the model architecture remains unconverged, similarities are high, mostly featuring a fusion of VLA, diffusion models, and world models, with layered reasoning (brain) and motor control (cerebellum) components. Participants are diverse: 1) major tech companies like Google and ByteDance; 2) embodied model and hardware companies such as PI, Figure, Zhiyuan, and Xinghai Tu; and 3) large language model firms like OpenAI.

Data is the core fuel for model training, with the key factors being data quantity, quality, and scale-up capability. Unlike large language models, which can leverage vast amounts of internet data for training, embodied intelligence naturally lacks robot trajectory data, making data a critical bottleneck for model iteration. Current data sources include internet data, simulation-generated synthetic data, first-person perspective data, and teleoperation data. Different companies are betting on various data pathways, generally adopting a mix of multiple data types and combining open-source with internal datasets. During the model training phase, opportunities arise for data producers and data collection equipment providers.

Risk factors include the risk of model technology iteration falling short of expectations, data supply and quality bottlenecks, and risks related to the commercialization of the technology.