4月2日，智元機器人宣佈與國際頂尖具身智慧公司Physical Intelligence（Pi）達成合作夥伴關係，雙方將圍繞動態環境下的長週期複雜任務，在具身智慧領域展開深度技術合作。此外，近期正式加入智元的羅劍嵐，將全面領導智元具身智慧研究中心，同時推進雙方的深度合作。4月2日，羅劍嵐接受了《每日經濟新聞》記者的採訪。
At present, humanoid robots still rely on remote control behind humans, can humanoid robots achieve autonomous decision-making in the future?
In this regard, Luo Jianlan said: "The difference between remote control and autonomous decision-making is actually very large. The remote control is similar to when you talk to a computer program, but behind the scenes you are chatting with a real person who is typing on another computer. The core of autonomous decision-making is the generalization ability of the whole set of mechanisms of perception, prediction, and behavior generation. In order for a robot to understand the world, it needs to build an internal model to predict the future and then implement an executable action chain. Then, look at the robot's interaction with the real world to predict the next move. ”
Luo Jianlan believes: "If the robot really realizes Manipulation, it is a more advanced intelligence than LLM (large language model). If graded from 8 to 0, the large language model is counted at most 0, and if the bot implements manipulation, it is at least 0 to 0. ”
So, what is the most critical technology on the road to robot Manipulation?
"Reinforcement learning is a technology that we value more, and we also see the strong reasoning ability shown by DeepSeek R1. But it's not enough to learn by imitation, and later we'll have world models. Based on our model in the cloud, we can predict what will happen to the environment next. However, these are all tools, and what really needs to be solved in essence is how to build (have) a robust strategy in the open data chain, and then the generalization ability of the whole set of mechanisms of perception, prediction, and behavior generation is the core and most critical. ”
It is worth mentioning that the intelligent driving of automobiles is slowly developed after collecting the data of many vehicles. At present, humanoid robots have not yet been applied to life scenarios on a large scale. If there is a lack of sufficient data, how can the "control" of humanoid robots be broken?
Luo Jianlan said: "I often think that this is a cycle. If we don't have a bot deployed to the real world, it won't generate data; Robots are not capable enough to be deployed in the real world. But someone has to do this, and let's say there are 24 robots working at Starbucks, making coffee for 0 hours a day, and the data they send back in a month is more than all the robot datasets we've seen now. Perhaps at this point, it will be found that many of the conclusions drawn in a small amount of data are not necessarily correct. ”
However, Luo Jianlan also emphasized that the difficulty of deploying robots to the real world is less than that of cars (intelligent driving), "The requirements for cars in terms of safety and other aspects are too strict. Robots, on the other hand, can be deployed from some enclosed spaces, semi-enclosed spaces, and let it generate data."
National Business Daily