Summary of Key Points
The prices of humanoid robots have recently plummeted significantly (from millions to tens of thousands of yuan), with the localization rate exceeding 90% and a high global market share. However, the actual adoption rate is less than 2%. The industry consensus is that robots do not need to resemble humans; practical value is more important than appearance. Currently, robots have begun to be implemented in fixed scenarios such as factories and warehouses, but large-scale industrialization still faces challenges related to technology, data, and cost. The primary technical approach is VLA (Vision-Language-Action), with global models gradually being integrated. Companies should prioritize practical robots with clear tasks and defined input-output relationships, rather than chasing the gimmick of having a humanoid form.
1. Humanoid Robot Prices Plummet, but Few Are Actually Used
Engineering prototypes that cost nearly a million yuan a year are now available for sale in the second-hand market for just 50,000 yuan each, and consumer-grade products are even cheaper than high-end iPhones (for example, Songyan Power's Bumi costs only 9,998 yuan). China’s supply chain has a localization rate of over 90%, and by 2025, 90% of global humanoid robots will be produced in China. However, the reality behind this excitement is sobering: According to Gartner research, only 1.64% of customers have actually deployed humanoid robots, while 98% are still in the exploration phase. Experts argue that robots do not need to look like humans; for instance, Amazon’s Digit has knees that bend backward, making it more efficient for squatting, and the 1X Eve uses a wheeled chassis for faster movement indoors. The human form is merely a gimmick; the ability to perform tasks effectively is what matters.
2. The Core of Robot Adoption: “Can They Generate Stable Profit in Fixed Scenarios?”
Robots are more likely to be successfully implemented in scenarios with clear task boundaries, repeatable processes, and few exceptions. Examples include logistics tasks on factory lines (moving parts) and warehouse handling (organizing goods). These environments are stable, reducing the likelihood of errors and making it easier for companies to calculate the cost-effectiveness of robot adoption. Home scenarios, on the other hand, are more challenging: tasks are dispersed (sweeping floors, cooking, picking up items), the environment is variable (toys may appear unexpectedly on the floor), and safety requirements are high (robots must not injure people). Therefore, home robots require more mature technology.
3. The “Brain” of Robots: VLA as the Main Technology, with Global Models Providing Assistance
The “brain” of a robot is its model, and the most advanced approach currently is VLA (Vision-Language-Action): vision is used to understand the environment, language to interpret commands (e.g., “Turn on the light because the room is dark”), and actions are taken to complete tasks. This approach is the opposite of previous techniques, which focused on perfecting a single action before learning others. Now, robots first develop generalization capabilities (understanding various commands) and then improve reliability in specific scenarios. Global models represent a new direction, enabling robots to directly predict physical phenomena (e.g., slowing down upon seeing a puddle without needing to anticipate potential slippage), but they are mainly used for simulation at present and have not yet been widely applied in real-world robots. In the future, these two technologies will merge to make robots smarter.
4. Two Barriers to Mass Production: Dexterous Hands and the Data Gap
1. Dexterous Hands: To perform precise tasks (e.g., grasping a cup), robots need sufficient freedom of movement while also balancing cost and durability. High-end dexterous hands from overseas cost tens of thousands to hundreds of thousands of yuan, which are too expensive; those priced in the thousands of yuan are not durable enough to replace human labor.
2. Data Gap: Training robots requires real-world operation data (e.g., remote control operations), but this is costly. Simulation data is cheaper, yet it differs from real-world conditions (perfections in simulation may fail in reality due to friction). Human video data (e.g., cooking videos) is also difficult to use directly because the structure of robot and human hands is different. A solution is to combine real data with simulation and human data.
5. The Window for Industrialization: Opportunities Lie in Practical Robots
Robots are on the verge of mass production. Capital is flowing into the industry, but large-scale adoption has not yet occurred (for example, Tesla’s Optimus will only perform basic tasks by 2025 and start mass production in 2026). In the short term, humanoid robots will only be used in limited pilots, while practical robots such as industrial robotic arms, warehouse AMRs, and service robots have a clearer path to market. Here are some recommendations for companies:
- Do not purchase humanoid robots immediately; focus on tasks with high value and low complexity (e.g., warehouse logistics).
- Treat robot implementation as an operational transformation project, involving more than just purchasing hardware; also adjust processes and layouts.
- Start with small-scale pilots before expanding to larger scenarios.
- Prioritize mature categories of robots (e.g., robotic arms) and wait for humanoid robots to become fully developed.
Conclusion: The Competition in the Industry Is About “Usefulness,” Not “Human-Like Appearance”
The essence of robots is as tools for enhancing productivity in the physical world. The significant drop in prices and the maturation of the supply chain are positive developments, but don’t let the gimmick of having a humanoid form distract you from the more important question: What practical problems can this robot solve? Practicality and value are more crucial than appearance. This race for productivity has just begun.