Summary of Key Points
This year, the humanoid robotics industry has seen two significant changes: First, there has been a surge in shipments (over 40,000 units in the first half of the year in China, accounting for 97% of the global total), although we are still far from large-scale commercialization. Second, the focus of the industry has shifted from showcasing technical prowess (such as running, jumping, and playing tennis) to practical applications (the ability to perform tasks effectively). The competition is no longer about athletic performance but about the efficiency with which robots can complete tasks in unfamiliar environments. At the same time, industry experts are defining the criteria for the "ChatGPT moment" of embodied intelligence—when robots can complete 80% of tasks using verbal commands—and believe that generalization ability and data accumulation are the biggest challenges at present. The business model is also evolving from selling robots to selling their productivity.
From “Showcasing Skills” to “Working as Workers”: The Practical Shift in the Robotics Industry
In recent years, humanoid robots have primarily demonstrated impressive physical abilities, such as YuShu's new robot running at 12.65 meters per second (faster than the fastest human), and Galaxy General's robot playing tennis. However, these feats are far from practical applications. Factories need robots to assemble parts, households need help with cleaning, and warehouses require sorting and packaging goods. This year, the industry has begun to focus on more tangible tasks. For example, YuShu demonstrated a robot that can clean a meeting room (switching between 7-8 tasks while handling interruptions), and the robot from StarSeaMap in collaboration with JD.com completed over 100 orders in a day without human intervention. Although these tasks may not seem exciting, Wang Xingxing believes they have practical value and are a step towards integrating robots into everyday life.
Behind the 40,000 Shipments: What’s Still Missing for Large-Scale Adoption?
The 40,000 units shipped in the first half of the year may seem substantial, but most were small-scale trials rather than widespread adoption. There are three main reasons for this:
1. Efficiency: Robots are not as efficient as humans. For instance, it takes a robot 2 minutes to fold a short-sleeve shirt, which is only 70-80% as fast as a human. Even in factories where tasks are simple, robots still perform slower than workers.
2. Lack of Generalization: Robots struggle to learn new tasks efficiently. It currently takes about 10 hours of training to teach a robot a new task, and their success rate drops significantly when the environment changes (e.g., if the location of items is altered). This is similar to an employee who can only perform tasks at one specific station and needs retraining for different roles.
3. Insufficient Precision: Robots can perform tasks with centimeter-level accuracy, but finer tasks (such as precise assembly) require further improvement through reinforcement learning.
These issues limit robots to specific use cases and prevent them from being widely applied in more complex scenarios.
The Business Model is Evolving: From Selling Robots to Selling Productivity
Robotic companies used to earn profits by selling the machines themselves (with gross margins of 40-60%). Now, they are shifting to a new model:
- Stage 1: Selling the robots outright.
- Stage 2: Offering subscription-based services, where fees are based on the number of tasks completed by the robot (gross margins around 20%).
- Stage 3: Creating a token economy in the physical world, where payments are made for the value generated by the robot—e.g., charging per task completed. The focus shifts from the hardware to the intelligent capabilities of the robot.
StarSeaMap has already delivered orders for thousands of robots and could reach tens of thousands next year. In the future, companies will be evaluated not based on the number of robots sold but on how much work they can perform, the efficiency of their training processes, and the success rate of their tasks.
How Close Are We to the “ChatGPT Moment”?
Last year, people were wondering when the era of embodied intelligent ChatGPT would arrive. This year, industry experts have provided specific criteria:
- Wang Xingxing (YuShu): The breakthrough will occur when robots can complete 80% of tasks in unfamiliar environments using verbal commands. This could happen in 2-3 years at the earliest or 5-10 years at the latest.
- Wang He (Galaxy General): Current embodied models are comparable to GPT2, but to reach the ChatGPT level, robots need to achieve a 70-80% success rate on common tasks and allow ordinary people to quickly train them for specific tasks. The company aims to achieve this by 2028.
The main challenges include:
- Generalization: Robots perform well in trained scenarios but fail when the environment changes.
- Lack of Data: Training robots requires millions of hours of data, but we currently only have tens of thousands.
- Misalignment between AI and Robotics: Robots understand the task (e.g., picking up an object), but minor errors (in vision or touch) can lead to failure.
The Future Focus: Data and Models Become the Core
In the past, robotics competition focused on hardware (speed, flexibility of joints). Now, the focus is on software and data:
- YuShu: Developing self-evolving physical AI models that automatically generate control code and continuously improve through simulation and real-world testing.
- All Companies Compete for Data: Data is essential for training robots, as it serves as the “training material” for their “brains.” Those with more real-world data will have a better chance of overcoming generalization issues.
In short, for robots to move from being laboratory curiosities to practical tools, the key is not speed but the ability to work stably in changing environments. This marks the true “coming-of-age” of the robotics industry.
This news indicates that the robotics industry is transitioning from a phase of showcasing technical feats to one where robots become reliable and effective workers. The focus in the coming years will be on developing robots that can truly contribute to business processes.