The idea is to provide customers with a complete technology stack from the chip level (Jensen Source family of chips), to the basic models (Project Groot), to the simulation tools and other utilities developed along the way. This will become the computing platform for humanoid robots and intelligent robots. I would like to quote Jensen, one of my favorites: “Everything that moves eventually becomes autonomous.Although it's not there yet, we can predict that in the next decade or so, if we believe that there will be as many smart robots in the future as there are iPhones, we better start building today. Sonia Huang, great! Are there any results from your research so far that deserve special mention? Is there anything that makes you feel confident or optimistic about your approach? Jim Fan Yes, we can talk about some of the previous work. One of the works that I'm really happy with is called URAC.
to train a five-fingered robot hand to perform the bolivia phone numbers movements of turning a pen. For me personally, this is especially funny because I gave up the skill of turning a pen. So I can't do it myself, but the robot hand can. And the method that we use to train it is to use LLM to write code that controls a simulation API that Nvidia built, called the i6M API. The LLM outputs code to define a reward function. A reward function is basically a specification of the desired behavior that we want the robot to perform. If the robot is on the right track, it is rewarded; if it does something wrong, it is punished. Typically, a reward function is designed by a robotics expert who is very familiar with the API, which is a job that requires a high degree of expertise, and the process is very tedious and manual.
We developed an algorithm that uses LLM to automate the design of reward functions, allowing robots to complete complex tasks such as turning a pen. This is a general technique and we plan to extend it beyond just turning a pen, it should be able to design reward functions for different tasks, and even generate new tasks via Nvidia's simulation API. This gives us a lot of room for further development. Sonia Huang I remember five years ago there were research teams that were solving the Rubik's Cube using robotic arms. In the last year or so, this field seems to be heating up again. Why do you think now is the “time” for robotics? Is there anything different? We’ve heard that OpenAI is also re-entering the robotics field and everyone is ramping up their efforts. Do you think anything has changed? Jim Fan Yes, I think there are a few key factors that are different now than before.