Jensen Huang's eldest daughter made an appearance on a live stream and talked about embodied intelligence.

People have seen a lot of Jensen Huang, but have you ever seen his daughter talk about embodied intelligence?

Well, Jensen Huang's daughter, Madison Huang, made her first public appearance on a live interview show. As the Senior Director of NVIDIA Omniverse and Physical AI, she had an in - depth discussion with Xie Chen, the CEO of Lightwheel Intelligence, and Mustafa, the Growth Lead of Lightwheel Intelligence, on "how to narrow the gap between virtual and real robots".

Lightwheel Intelligence is a company focusing on simulation synthetic data technology. Different from companies focusing on large - scale models, their core goal is to help AI better understand and enter the physical world. Currently, they mainly focus on two scenarios: embodied intelligence and autonomous driving.

During the one - and - a - half - hour interview, the three put forward a series of important viewpoints:

Synthetic data is crucial for solving the robot data dilemma.

The SimReady assets of Lightwheel Intelligence not only need to be visually accurate but, more importantly, physically accurate.

NVIDIA and Lightwheel Intelligence are jointly developing Isaac Lab Arena - a next - generation open - source framework and platform for benchmarking, evaluation, data collection, and large - scale reinforcement learning.

...

Let's take a closer look below.

Using synthetic data and simulation to solve robot data obstacles

As Interview 1 officially started, the host, Edmar Mendizabal (Omniverse Community Manager), straightforwardly raised a question that many people are curious about.

How did the cooperation between NVIDIA and Lightwheel Intelligence start?

Madison explained that many internal projects at NVIDIA rely on the support of Lightwheel Intelligence. For example, Gear Lab is building a general agent model, and the Seattle Robotics Lab is carrying out a large number of tasks involving contact operations and precision assembly.

For language model researchers, they can use data from the entire Internet to train LLMs. However, in the field of robotics, the situation is completely different. They have to manually collect data, which is why so many data collection factories have emerged.

In this data - scarce situation, NVIDIA believes that simulation is the solution. Therefore, a synthetic data factory is needed, and it also hopes that partners share the vision of OpenUSD and use it as the basis for building SimReady Assets.

In 2023, Lightwheel Intelligence was founded with the goal of using synthetic data and simulation to break through the robot data bottleneck.

But at that time, the field of robotics was still in a very early stage, so they started with the synthetic data problem in autonomous driving. Subsequently, the cooperation almost expanded to all teams at NVIDIA.

Interestingly, Xie Chen used to be the person in charge of NVIDIA's autonomous driving simulation. After going through a roundabout way, he is now working for NVIDIA again.

Next, the host asked, "What problems still exist in the transition from virtual to real (Sim2Real) for robots now?"

Xie Chen replied:

For autonomous driving, Sim2Real is the easiest to solve because it mainly relies on visual perception. For robots, everything involves physical contact, and the most important thing is the manipulation ability. At the same time, it also requires the cooperation of dexterous hands and tactile sensors, so the problem becomes more complicated.

The core problem lies in physical accuracy.

Take a refrigerator as an example. When you open the door, you can feel the force of the magnetic sealing strip, and when you pull the drawer, you can feel multiple frictions. These physical characteristics are very precise.

To achieve this physical accuracy, data is very important. High - quality data is the key to entering the robot training system and generating correct algorithms.

Therefore, Xie Chen also specifically mentioned the concept of the digital pyramid.

He believes that deploying embodied intelligence in the real world requires a huge amount of data, actually more than that required by large - scale language models. This creates a huge data obstacle, and real - world data cannot completely solve this problem.

Take autonomous driving as an example. There are a large number of drivers and cars on the road in reality, but in environments such as factories and homes, the number of robots is very limited.

Therefore, synthetic data will become the most important and main data source for solving the data bottleneck of embodied intelligence.

They used a large number of physical devices to collect accurate data and implemented it in the simulation environment. At the same time, they also designed some ways to compare the forces in the real world with those in the simulation to ensure they match.

Besides data, another point that Xie Chen considers important is efficiency.

He mentioned that reinforcement learning is very important, but to run large - scale reinforcement learning, it is necessary to ensure that different types of simulations are very computationally efficient.

To run a large number of simulation environments simultaneously, they use simple and efficient methods (such as basic geometric shapes and convex hulls) to detect collisions, which can not only maintain sufficient accuracy but also save a large amount of computing resources.

After that, Xie Chen also talked about cable simulation. Cables are like flexible objects but also behave like rigid bodies in some cases, so their simulation is actually very difficult.

To enable robots to learn how to operate cables, Lightwheel Intelligence cooperated with Newton and NVIDIA to build a solver for cables and developed simulation - ready assets to build this simulation.

As we all know, the difference between humans and animals is that humans can use tools. So, how to teach robots to correctly use tools to complete specific operations is becoming increasingly crucial.

For example, it is very difficult to let a robot cut a cucumber in the simulation. This is not only for data collection but, more importantly, to support reinforcement learning.

For this reason, Lightwheel Intelligence has cooperated with NVIDIA's Isaac Sim Lab to jointly tackle the challenge of transferring from simulation to reality.

Finally, Xie Chen mentioned that Lightwheel Intelligence is also jointly building Isaac Lab Arena with NVIDIA - a framework platform for next - generation benchmarking, evaluation data collection, and large - scale reinforcement learning. This project was officially launched by NVIDIA at the CoRL conference.

Jensen Huang's children

After the interview, let's dig into Jensen Huang's two rarely - seen children.

First is his daughter, Madison, whose Chinese name is Huang Minshan and is 34 years old.

In 2020, she joined NVIDIA as a marketing intern. After four months of internship, she became the event marketing manager of the Omniverse department and has been working in this department ever since.

Madison has successively held positions such as product marketing manager and senior product marketing manager at NVIDIA until she became the senior director in March this year.

Surprisingly, Madison's initial career was in cooking.

In 2012, she obtained a Bachelor of Business Administration in Culinary Arts from the Culinary Institute of America. Then she went to Le Cordon Bleu to learn dessert - making and wine knowledge and worked as a chef in New York and San Francisco.

In 2015, Madison returned to Paris and joined the luxury goods industry, working as a marketing and development manager at LVMH. During her time at LV, Madison also took a short - term course in data science at the London School of Economics.

In 2019, Madison and her brother, Spencer, took a short - term AI executive course at MIT.

After that, she obtained an MBA degree from the London Business School in 2021 when she was already a full - time employee at NVIDIA.

After talking about Jensen Huang's daughter, how can we not mention his son?

Similarly "following in his father's footsteps" is Madison's brother, Spencer, whose Chinese name is Huang Shengbin and is 35 years old this year.

His position at NVIDIA is the robotics product line manager, responsible for developing AI models and simulation software for robots.

Spencer joined NVIDIA in 2022 and initially served as the product manager of the Isaac Sim Cloud team.

As mentioned when introducing Madison, the siblings once participated in the short - term AI executive course at MIT together. However, Spencer also took an additional course on human - computer interaction.

After that, Spencer first took a short - term course at Harvard Business School and then an MBA at New York University, obtaining the degree in 2022.

Interestingly, even earlier, Spencer was a bar owner.

In 2012, Spencer graduated from Columbia College Chicago, the largest private art and media college in the United States, majoring in international marketing and cultural studies.

After graduation, Jensen Huang asked him to go back to "his hometown" to learn Chinese for a year. During this period, Spencer founded his cocktail bar, R & D Cocktail Lab, and ran it for eight years.

It is reported that this bar has won many international awards and was once selected as one of the top 50 bars in Asia. However, Google Maps currently shows that the bar has permanently closed.

Well, well, well, the rich second - generation is going to focus on inheriting the family business, right?

Reference link:

https://www.youtube.com/watch?v=UgT - P6ynxLc

This article is from the WeChat official account "QbitAI", author: Shiling. It is published by 36Kr with authorization.