News Analysis in Plain Language
Recently, there's been a significant shift in the AI community: while previously the focus was on the flexibility of joints and the power of motors (the “physical aspects” of robots), now the entire industry has turned its attention to the “brains” of robots. The catalyst was OpenAI’s announcement of its plans to enter the humanoid robot market, followed by the release of GPT-6 Astra, which revealed some limitations. GPT-6 Astra had a 95% success rate in simple tasks like picking up and placing objects, but its success rate dropped to 10% when it came to more complex tasks, such as plugging a plug into a socket. This indicates that even the most advanced large models struggle when faced with the real-world challenges of physical interactions.
The industry has realized that while these models excel at creating content and generating images in the digital world, they are utterly inadequate when dealing with the unpredictable and complex realities of the physical world. Investments have already begun to reflect this shift—over the first half of the year, more than 40 billion yuan was invested in the field of embodied intelligence in China, with over half of that going to companies developing “robot brains,” particularly those focused on so-called “world models.” However, a recent independent test of 12 open-source embodied models showed that their highest success rate for complex tasks was only 16.7%. There's no consensus among the industry about the actual capabilities of these world models.
To address these issues, industry leaders, investors, and university professors gathered for a closed-door meeting to discuss four key questions: Are world models truly essential for robots to function effectively? What constitutes a “real brain” that can perform tasks? Are the current high valuations based on genuine technology or just inflated expectations? And which technologies will be able to make a real impact in the coming year? On September 22nd, a live event will be held for all industry participants to exchange resources, setting a benchmark for evaluating the progress of embodied intelligence one year from now.
---
Detailed Explanation of the Key Points
1. The Real Challenge for Humanoid Robots: The Brain Is 100 Times More Critical Than the Body
Many people thought that as long as humanoid robots had joints as flexible as humans, they could be used in factories and for household chores. OpenAI’s tests shattered this illusion. Even if a robot’s joints are more precise than human joints, if its brain cannot predict the movement of objects with enough accuracy, it cannot perform complex tasks. Previous large models were trained solely in the digital world and had no experience with real-world obstacles—things like slippery cups, deformable bread, or small stones on the ground. Now that the hardware of humanoid robots has reached a satisfactory level, the “intelligence” of their brains is still akin to that of a 3-year-old child. This has led the industry to focus on developing more advanced brains, as the previous approach is no longer viable.
2. The So-called “World Models” Aim to Give Robots Human-Level Predictive Abilities
When asked about world models, different companies give different definitions. Some claim they can generate images that robots will see next; others say they can create simulated training scenarios without the need for extensive experimentation; still others claim they can predict whether a robot will damage objects while trying to pick them up. In essence, the goal is to impart the physical common sense that humans have acquired over decades directly into robots. For example, a robot should be able to anticipate the consequences of its actions before they happen, such as knowing it will knock over a cup before it even touches it. However, most current world models are still at a basic level of generating realistic simulations and do not truly understand the laws of physics.
3. Capital Is Betting on the Next Monopoly
Over the first half of the year, more than half of the 40 billion yuan in investments in embodied intelligence went to brain development. The logic behind this is simple: if humanoid robots become widespread, it’s unlikely that each robot will have its own brain system developed from scratch. A universal, intelligent system that all robots can use will emerge, and the company that develops it will gain a huge advantage, similar to the dominance of Android and NVIDIA in the smartphone market. However, there’s also a significant risk of over-inflated valuations, as many companies rely on speculative claims rather than concrete data. The closed-door meeting aims to clarify the distinction between genuine technology and hype.
4. Who Will Control Robots in the Future?
There are two opposing views in the industry: companies that build the physical bodies of robots argue that they know best how their joints work and should control the intelligence; those that develop the brains argue that the cost of developing universal models is too high. It’s likely that the industry will follow the path of smart cars, where a few large companies will develop their own brains, while most smaller manufacturers will use third-party solutions. The ultimate winner will be the one with the most real-world data and experience. The ability of a brain to improve a robot’s performance significantly and to adapt quickly will be the decisive factor.
5. How to Tell If a Robot Brain Is Truly Mature
To determine if a robot brain is mature, look at these six key indicators:
- The success rate of completing a task repeatedly (not just a single successful attempt).
- The ability to adapt to changes in the environment (e.g., moving objects, changing lighting, or adding obstacles).
- The ability to recover from mistakes and continue working autonomously.
- The time and cost required to adjust the system for new tasks.
- The time and cost needed to transfer the brain to a different robot model.
- The frequency of human intervention required during tasks.
Robots with high scores in these indicators will be ready for practical use in factories and homes, and the industry will no longer rely on continuous funding.