Summary of Key Points
The current humanoid robotics industry is at a stage where there is a disconnect between hype and reality: Robots performing impressive tasks such as dancing and folding clothes at exhibitions make people look forward to the arrival of a era of general-purpose robots. However, in real-world applications, robots are far from capable of as much as they demonstrate. The core issue lies in the insufficient coordination between the “brain” (embodied intelligence) and the “body” (hardware). The brain can issue commands, but the robots often have difficulty grasping objects with precision, and their technical approaches are not yet unified, leading to a lack of relevant data. Although there is a bubble (where valuations far exceed actual value), in the long run, the industry is expected to experience a breakthrough similar to that of ChatGPT once these data and value issues are resolved, likely within 2 to 10 years.
1. Exciting Displays, but Incompetent Robots
Robots that can dance and fold clothes at exhibitions may seem intelligent, but in factories or homes, they fail to perform basic tasks effectively. For example, they may slip while trying to grab a cup, crush a box due to excessive force, or be unable to open a locked cabinet. The reason is simple: the brain can plan the actions, but the body cannot execute them properly.
The main limitations of robots’ physical capabilities lie in three areas:
1. Incompetent hands: The control of strength and precision in dexterous hands is inadequate, resulting in either poor grip or damage to objects.
2. Unstable feet: Two-legged robots often stumble and lack balance.
3. Lack of coordination: Movements are not smooth, leading to pauses when performing tasks like picking up an object from a table and placing it in a cabinet.
Industrial applications require much higher precision—millimeters or even sub-millimeters, which current robots (with precision in the range of centimeters to millimeters) cannot meet. As a result, robots are mainly used in research, performances, and education, with industrial applications being more of a novelty than a practical solution.
2. Three Major Barriers to Practical Adoption
Even if robots could perform basic tasks, there are three significant hurdles to their widespread use in factories and homes:
1. Lack of generalization: Robots perform well in specific, controlled environments but struggle in new situations.
2. Insufficient precision: Industrial tasks require sub-millimeter precision, which current robots cannot achieve.
3. Low efficiency: Robots are much slower than humans in simple tasks, making them less attractive to manufacturers.
There is also a deeper issue: The physical world is too complex for robots to handle all situations with a single set of rules. Models and data cannot account for unexpected events, such as a part falling in a factory or a child dropping a toy at home.
3. Technological Progress: From Imitation to Predictive Learning
Robots’ “brains” (embodied intelligence) are undergoing a major upgrade:
- Past: VLA models (imitation learning): These taught robots to mimic human actions by showing them numerous videos. However, they could only work in static, fixed environments.
- Current: World models (predictive learning): Robots can predict future events, such as whether a cup will slip or whether the ground has uneven surfaces, and adjust their actions accordingly. This addresses the limitations of imitation learning but is still in the laboratory stage.
- Common challenge: The lack of data is a major bottleneck. Current embodied models have only around 7 billion parameters (compared to the hundreds of billions in language models), and the training requires fewer GPUs (a few dozen cards are sufficient). To achieve the level of intelligence seen in ChatGPT, thousands or even tens of thousands of GPUs and more real-world data are needed.
4. A Bubble, but a Rational Industry Growth
The industry currently has a bubble, with companies valuing themselves significantly higher than their actual worth. For example, Yuzhu Technology, the first humanoid robotics company to go public, saw its stock price halve because the market realized its products had not yet been successfully deployed in real-world applications.
However, a moderate bubble can be beneficial, as it attracts investment and talent, accelerating technological advancement. Industry experts believe that two key factors will drive the industry’s future breakthrough:
1. Solving the data shortage: Collecting more data from real-world scenarios (e.g., videos of robots working in factories and homes).
2. Creating practical value: Robots should be able to perform tasks that humans find difficult or impossible, such as working in mines or caring for the elderly.
When will we see a true breakthrough? Wang Xingxing, the founder of Yuzhu Technology, believes it will happen when robots can perform about 80% of tasks in any unfamiliar environment. This could take 2 to 3 years at the earliest or 5 to 10 years at the latest.
The lights at the exhibitions may be bright, but the real test lies in the less visible areas—factories and home kitchens, where robots need to work reliably day in and day out.
In conclusion: The humanoid robotics industry is promising but faces significant challenges in terms of hardware, data, and technology. Only by overcoming these obstacles can we move from impressive demonstrations to practical applications. The bubble will eventually burst, leaving behind companies that have truly solved the problems that robots are designed to address.