Summary in Plain Language
The field of "embodied intelligent robots" – robots that can independently interact with the physical world, perform tasks, and move around – has become incredibly popular recently, with high levels of funding and media attention. However, the entire industry previously followed the lead of ChatGPT, believing that simply accumulating data and parameters would lead to rapid progress, which led to many detours. At the recent Bund Conference, several top entrepreneurs and scientists in the industry discussed the latest consensus: the development path of embodied intelligent robots cannot be replicated from that of large language models. Questions such as "how much real data to collect," "whether general-purpose models will compete with robots for jobs," and "when robots will become profitable" now have clear answers. The future competition will not be about which model is the most powerful, but rather about the comprehensive capabilities of data, hardware, systems, and scenarios. The industry is transitioning from a focus on hype and demonstrating concepts to practical applications that generate real revenue.
---
Detailed Analysis
1. The industry has just realized its first major mistake: thinking that copying ChatGPT’s approach of accumulating data would work
Large language models improved significantly by being fed trillions of words of internet text. Consequently, many assumed that embodied intelligent robots would also improve simply by collecting millions of hours of video data showing them performing tasks. However, it has become clear that this approach is fundamentally flawed. Most of the data collected is useless or of poor quality. For example, training a robot to repeatedly lift a bottle of mineral water won’t teach it how to handle a thermos cup or a kettle with hot water. The consensus now is that the "quality" of data is far more important than the "quantity." Data that reflects real-world physical principles—such as how to grasp objects when they slip, how to avoid falling on wet surfaces, and how to handle soft objects without damage—is far more valuable. The ineffective data collected by many companies in the early stages of the industry was essentially a necessary learning cost.
2. The approach to data collection has completely changed: no longer competing on the amount of data, but on the ability to replicate real-world conditions in simulations
There was a long debate about whether to use data from real-world scenarios or data generated in computer simulations for robot training. The new direction is to directly replicate the physical characteristics of the real world in simulations. For instance, by accurately inputting details like the shape of water stains on tiles, the friction of the floor, and how different objects affect the stains, robots can practice countless times without wasting resources or damaging themselves. This approach is hundreds of times more efficient than traditional methods.
3. The fear of GPT models affecting the robot industry is unfounded
OpenAI’s GPT-6 Astra, which can directly control robots, sparked concerns that general-purpose models would threaten embodied intelligent startups. However, this concern is unfounded. AI capabilities are derived from specific data, and general-purpose models currently lack the necessary physical data for robot training. Creating such models would require starting from scratch, collecting sensor data, adapting to various hardware, and building entire simulation systems, which is just as challenging as for startups. In fact, startups that have already overcome many hardware and practical challenges possess unique data that general models lack, making their models more practical for real-world applications.
4. The criteria for robot commercialization are simple: whether customers are willing to pay
Previously, robot demonstrations focused on impressive tasks like serving tea, folding clothes, or tightening screws, but in reality, these robots often failed in new contexts. The new standard for commercial success is whether customers are willing to continue paying for the robots’ services. A robot is considered successful if it is more cost-effective and effective than traditional automation solutions, with a positive cost-benefit ratio. The focus is no longer on finding "killer applications"; any task that can be performed efficiently and reliably in real-world scenarios, such as delivering food in restaurants or assisting in nursing homes, can be profitable. Experts predict that truly commercialized embodied intelligent robots will emerge within three years, and the transformation of the physical world by AI is just beginning.