Summary of Key Points
This news article focuses on the industry developments regarding robots and embodied intelligence at the Yabuli Innovation Annual Conference: Currently, the large models for embodied intelligence (the “smart brains” of robots) are still in their early stages—similar to the period 1-3 years before the release of ChatGPT. The technological approach is shifting from simply “recognizing pixels” to developing models that can understand the underlying causal relationships in the world. In terms of implementation, industrial applications are making progress faster than consumer (e.g., home) scenarios. 2026 is expected to be the year of mass production, and by next year, robots that can perform tasks reliably may become a reality. However, the industry still faces challenges such as high hardware costs, limited computing power, and poor generalization abilities. Companies are actively investing in research and development (R&D) and seeking funding.
1. Large Models for Embodied Intelligence: Still in the Learning Phase, How Close to the “ChatGPT Moment”?
You can think of large models for embodied intelligence as the “brains” of robots—these brains need to be able to understand instructions, perceive their environment, and make decisions to complete tasks. At present, these brains are not yet sophisticated enough; according to Chen Li from Yushu Technology, they are at a level equivalent to that 1-3 years before ChatGPT was released and are still unable to handle most unfamiliar situations.
When will they be considered mature? Chen Li provided a clear benchmark: robots should be able to complete 80% of tasks using voice or text commands in 80% of unseen scenarios. For example, if you ask a robot to retrieve documents from an unfamiliar office, it would need to find the door on its own, avoid obstacles, and identify the documents—something it currently cannot do. Therefore, the widespread adoption of robots is hindered by the immaturity of their “brains.”
2. Technological Shift: From Pixel-Level Replication to Understanding the World’s Rules
Previous technologies for embodied intelligence (such as the VLA architecture) were more like trying to copy a cat’s appearance—robots relied on vision to recognize pixels and language to understand instructions, but they would get confused when faced with unfamiliar situations. The industry is now moving towards “hidden space world models,” which essentially teach robots common sense and causal relationships. For instance, Zhang Yufeng from Wujie Power explained that these models do not attempt to replicate every detail of the world (like the dust on a table); instead, they focus on capturing key principles (e.g., that pushing a cup will cause it to fall or that you need to turn a doorknob to open it). This approach allows robots to understand general rules, making them more adaptable to new situations. Although this technology is more challenging, it is considered the right path towards creating truly versatile robots.
3. Implementation Pace: Industrial Applications Take the Lead, Consumer Use Still Years Away
Industry experts generally believe that robots will first be used in industrial settings rather than in consumer households. Why? Industrial scenarios are relatively predictable and stable—robots can perform repetitive tasks like moving parts on assembly lines or organizing goods in warehouses, which are easier to train for. In contrast, home environments are much more chaotic (the location of sofas and the arrangement of cups can vary), making it difficult for robots to adapt.
Zhang Yufeng estimates that there will be 20,000 humanoid robots produced globally in 2025, but most of them will not be capable of performing meaningful tasks. 2026 is expected to mark the beginning of mass production, with robots achieving significant breakthroughs (e.g., being able to perform one or two industrial tasks reliably). Next year, we may see robots that can work consistently in factories without making mistakes, which could accelerate the pace of mass production.
4. Hurdles to the Widespread Adoption of Robots: Hardware, Computing Power, and Cost
Even if the “brains” of robots become more advanced, several practical issues need to be addressed for widespread adoption:
- Hardware: Robots need cheaper and more durable components. Currently, joints and sensors are expensive and have a short lifespan, making them unaffordable for most people.
- Computing Power: The brains of robots require substantial computing resources, but the computing power available on the robot itself (for battery life, heat dissipation, and cost) is limited. Additionally, there are latency issues with remote servers, which can affect robot responsiveness.
- Mass Production: Robots must be manufactured in large quantities, similar to how smartphones are produced, to reduce costs significantly.
Without solving these problems, robots will remain a niche product and not become a part of everyday life.
5. What Are Companies Doing?
Companies in the industry are actively making strategic moves:
- Yushu Technology: Focusing on developing the “brains” for embodied intelligence and collecting real-world data (from factories and offices) to train robots autonomously.
- Wujie Power: Working on hidden space world models and seeking funding to support R&D efforts.
- Other companies, such as Moushen Intelligence, are also exploring similar technologies.
Everyone is striving to gain a competitive advantage in this field, as the company that develops the most advanced “smart brains” will likely lead the robotics revolution.
In summary, embodied intelligence robots are currently in a preparatory phase before the dawn of widespread adoption. While technology is advancing and mass production is starting, there is still a long way to go before robots become commonplace. However, the industry is optimistic, and we may see reliably functional robots within the next year. Let’s stay tuned to see how this develops!