China's first robot with a truly meaningful multi-task continuous generalized embodied model was born! This robot is really trained by the model, it is understood that up to now, in addition to Figure 01, there seems to be no second company in China that can achieve this level of generalization ability, even if it is difficult in every way, it can complete the task. Tsinghua alumni started their own businesses and raised nearly 200 million yuan in just 4 months.

The just-concluded World Robot Conference has not dissipated its residual heat.

The flower work displayed by each robot is dazzling.

However, probably all the people who visit the exhibition have this feeling: at present, there are still very few robots in the world that really have strong generalization capabilities and can cope with various scenes.

Is there any robot in China that can achieve true multi-task continuous generalization capabilities?

There really is! Well-informed us learned that an embodied intelligence company called "Spirit AI" has demonstrated for the first time a powerful multi-task continuous generalization capability.

It is said that this mysterious company has not been around for a long time, but it has demonstrated such a level of technological advancement, how exactly did they do it?

Recently, we went to a factory and recorded in detail the jaw-dropping moments.

All actions are automatically generated by a neural network

No matter how difficult it is, he is not afraid, and his movements are extremely silky

With the cooperation of the researchers, we recorded a demo on the spot.

The little brother with the white paper cup approaches the robot and asks for "a cup of espresso", and while he is busy looking at the information on his phone, he accidentally knocks over the cup.

Let's see, what will Chihiro's robot do?

I saw that it was able to straighten the paper cup with one hand.

All actions are automatically generated by a neural network

Then place it on the coffee maker with your other hand and press the function key.

All actions are automatically generated by a neural network

When the coffee is ready, place the cup full of coffee in the middle of the table and you're done.

All actions are automatically generated by a neural network

The next guy who came over wanted a cappuccino.

But this time, it's a clear glass.

And just when the robot was about to reach the cup, the little brother deliberately "difficult" it and quickly removed the cup.

All actions are automatically generated by a neural network

In the face of this difficulty, the robot said that it was no problem at all!

The powerful generalization ability of the end-to-end neural network allows the robot to accurately identify transparent reflective objects, and its "clamp" hand can easily hold the cup no matter where it is placed.

Next, we came out and decided to give it a difficult mission.

Place a tissue box next to one of the robot's hands, then place the paper cup next to the tissue box, and ask for a cup of American.

Unexpectedly, it recognized the obstacle next to it at a glance, moved it aside, and successfully retrieved the cup.

All actions are automatically generated by a neural network

At the end, we drank the American style made by the robot.

What's even more surprising is that we also unexpectedly found on the spot: the "clip" hand of the Chihiro robot can also be replaced with a dexterous hand!

Not only does it pick up the apple with precision, but it can also turn it upside down and hold it with precision.

Moreover, it is said that no matter what kind of hand – two, three, or five fingers – can achieve continuous multitasking generalization.

After seeing this, we were amazed and could imagine the future of Chihiro robots entering the home, helping people complete a variety of tasks with their powerful generalization capabilities.

All actions are automatically generated by a neural network

Immediately afterward, we rushed to the conference room and asked curiously, "How exactly did the Chihiro robot achieve such a perfect continuous generalization ability?"

Luxury startup team, full stack AI

This embodied intelligence company, which can be called the "Chinese Figure", is also extremely scarce in the world.

The behind-the-scenes technical team has come from UC Berkeley, CMU, NTU, Tsinghua University, Peking University, Zhejiang University, Huawei, Tencent, DJI, Xiaomi and other top universities and enterprises at home and abroad.

In the research and development of embodied large models, robots, and practical implementation, this team that integrates academic elites and industry leaders has demonstrated outstanding strength.

They not only have core technical capabilities such as base model pre-training, RL, and IL, but also are at the forefront of the industry in terms of robotic arm system design, robot safety, and control architecture.

Because of this, Qianxun Intelligence is able to have full-stack AI engineering capabilities.

Let's get to know the key people in this star-studded leading team.

Founder and CEO Han Fengtao

Founder and CEO Han Fengtao, under the tutelage of academician Ding Han, an academic expert in robotics, has been deeply engaged in the field of robotics for more than ten years.

He was the co-founder &CTO of ROKAZ Robotics, the leader of high-performance light industrial robots in China and the first person in domestic force control collaborative mass production and delivery, leading the team to successfully deliver dozens of models of products with more than 20,000 units.

It is worth mentioning that these products have also obtained 43 domestic and foreign certifications, including the only two in the world and the only one in China for medical IEC60601 safety.

Moreover, the autonomy rate of the whole machine has reached more than 90%.

In terms of product application, Dr. Han Fengtao led the team to land the commercialization of 20+ industries, 100+ scenarios, and 1000+ customers.

In addition to his extensive practical experience, he has actively participated in many national research projects.

In February this year, Dr. Han Fengtao founded Spirit AI, which is committed to building industry-leading general-purpose robot AI systems and humanoid robots.

Chief Scientist Gao Yang

In terms of AI capabilities, we have to mention another core figure, Chief Scientist Gao Yang, who is also the co-founder of Chihiro Intelligence.

He studied in the Department of Computer Science of Tsinghua University under the tutelage of Professor Zhu Jun, a well-known scholar in the field of ML in China.

With his outstanding performance, he was awarded a full scholarship to the Department of Computer Science at UC Berkeley to pursue a PhD in Computer Vision.

During this period, Gao Yang studied under Professor Trevor Darrell, an international computer vision master (who has trained many well-known scholars in the field of vision, including Jia Yangqing).

In addition, during his Ph.D. and postdoctoral studies, he worked closely with leading scholars in the field of robotics learning, Sergey Levine and Professor Pieter Abbeel.

Pieter Abbeel is one of the proposers of the diffusion model (the core technology behind Sora and SD).

Aravind Srinivas, founder of Perplexity AI, a popular AI search startup, and John Schulman, a former OpenAI co-creator, are both his students.

Address: https://arxiv.org/pdf/2006.11239

In addition, Professor Sergey Levine is the founder of Physical Intelligence (Pi) in United States. Pi is a leading company in the field of embodied intelligence in United States, receiving a total of $70 million in angel investment from OpenAI and other companies.

During his Ph.D., Gao Yang published a paper on end-to-end autonomous driving with large-scale real-world data at CVPR, the top AI conference.

This laid the academic foundation for later end-to-end autonomous driving, such as FSD.

Address: https://www2.eecs.berkeley.edu/Pubs/TechRpts/2020/EECS-2020-5.pdf

In terms of personal contributions, all the core technologies behind the robots introduced above are all contributed by Chief Scientist Gao Yang.

He has achieved fruitful research results in the three-layer model of embodied intelligence.

In terms of reinforcement learning, Gao Yang proposed EfficientZero and EfficientZero v2, the world's most efficient reinforcement learning algorithms so far.

EfficientZero has been highly praised by John Schulman, a former OpenAI co-creator and leader of reinforcement learning.

In terms of imitation learning, he proposed the EfficientImitate high-performance imitation learning algorithm, which improves the performance by 600% compared with Stanford's VMAIL.

In addition, Gao Yang proposed ViLa and CoPa models from the perspective of using Internet video and pre-trained VLM.

In terms of hardware, Qianxun Intelligence is also a leader in this track.

The team not only has world-class robot motion control system development capabilities, but also has first-class system-level electromechanical system design capabilities

The most important thing is that they already have rich experience in industrial robots and medical robots, and they have achieved "dimensionality reduction" in technology.

In short, Qianxun's leading edge in software and hardware has also become a key factor in continuing to attract and win the favor and bet of capital.

All actions are automatically generated by a neural network

In 4 months, 200 million yuan was raised

The angel round of financing of Qianxun Intelligence Spirit AI was led by Honghui Fund, followed by Dachen Caizhi and Qiancheng Capital, and at the same time, Shunwei Capital and Oasis Capital continued to increase their weight as old shareholders of the seed round.

Now, the next flashpoint of embodied intelligence is just around the corner. In terms of commercial services and household services, it may be 3 to 5 years to land in batches.

From industry to services to home applications, a trillion-dollar market that needs to be explored urgently is unfolding in front of everyone's eyes.

With industry-leading embodied large-scale model technology and excellent robot R&D capabilities, Qianxun Intelligence will complete the commercial closed loop from technology R&D to product marketization with the highest efficiency.

General robots have become close partners of human beings, and they are about to move from science fiction to reality, and the whole world has begun to enter the era of intelligent robots. The moment when robots can be used by everyone like iPhones may be just around the corner!

After reading the robot making coffee autonomously, we have a more concrete understanding of the reasons why Qianxun Intelligence impresses investors.

In the view of the investment team of Honghui Fund, embodied intelligence is an important application scenario of AGI, and the market space is extremely broad.

In the past, the control of robots relied on a large number of manual programming processes, which had many limitations on the scene. The agent formed by the combination of embodied large model algorithm and hardware will greatly improve the task generalization compared with traditional robots.

This type of agent will be the best path to spatial intelligence. China has a leading edge in the robot hardware industry chain.

The team is very much looking forward to the mass production of a new generation of intelligent robots just around the corner, setting off a new industrial revolution.

The Shunwei investment team is very optimistic about the compound background and industry experience of the founding team of Qianxun Intelligence. Similarly, the head of Oasis Capital Investment also said that Qianxun's team not only combines industry understanding, but also accumulates cutting-edge technology. It is also because of this that we have witnessed the company's efficient growth and iteration in the short term.

The investment team of Dachen Caizhi affirmed the scarcity of Qianxun Intelligence in the industry - such a team is a veteran team with robot hardware, embodied AI algorithm capabilities and commercialization experience, which closely combines the past understanding of robot engineering and cutting-edge academic accumulation.

In just half a year, the company's embodied large model and rapid software and hardware iteration capabilities are all impressive.

The investment team of Qiancheng Capital expressed strong confidence in Qianxun Intelligence.

In their view, the revolutionary breakthrough of the AI model has given the robot more intelligence and agility, showing generalization and generalization capabilities.

In the next trillion-level track of humanoid robots, Qianxun Intelligence is undoubtedly at the forefront of the industry. 

This article is from Xinzhi self-media and does not represent the views and positions of Business Xinzhi.If there is any suspicion of infringement, please contact the administrator of the Business News Platform.Contact: system@shangyexinzhi.com