news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Zhidongxi (public account:zhidxcom）
authorvanilla
editLi Shuiqing

Zhidongxi reported on August 20 that yesterday, the startup Astribot released a new generation of AI robot assistant Astribot S1, which demonstrated the ability to perform a series of difficult, long-sequence, and generalizable tasks, including playing the piano, feeding cats, making pancakes, and making tea.

▲Robot S1 making tea

The robot will make its debut to the public at the Beijing World Robot Conference from August 21 to 25, starting tomorrow, and is expected to be delivered in batches in December this year.

Lai Jie, founder and CEO of Stardust Intelligence, shared the latest progress and technical path of the robot S1 with a few media outlets including Zhidongxi, and had in-depth exchanges with us on key issues such as commercialization progress, technological trends, and future plans.

When Zhidongxi asked, which specific customers or scenarios will the mass delivery of products in December target, and in what form will the delivery be?

Lai Jie said that the first batch of deliveries is mainly for scientific research scenarios, and both parties hope to reach a cooperative relationship rather than a transaction to jointly promote the development of the entire industry. He revealed that when Stardust Intelligence first released the demonstration video in April, it had actually received a lot of orders, but the team believed that at this time, its own technology and embodied industry are still strongly bound, and it needs to iterate products through cooperation.

Talking about future plans, he said that the team will have two core points next year. One is to focus on the technical advantage of "force" and continue to pursue human-like force perception and control in the underlying hardware design; the other is to pursue the ultimate in AI imitation of human beings, so that robots can quickly replicate the scarce abilities of humans.

1. Wheeled chassisThe human figure upper body, the "clip hand" accurately controls the strength

As can be seen from the demonstration video, the robot S1 can complete various complex tasks independently or under remote control, and its movements are very smooth and fast.

For example, S1 can help you take care of your cat when you are not at home, pour cat food for it, or help you play with the cat under VR remote control.

▲Robot S1 pours cat food and teases cats

It can also help you make pancakes, and it will use a plate to "flip the spoon" to turn the pancakes over.

▲Robot S1 making pancakes

In terms of whole-body coordinated control, the S1 can bend by coordinating the lower limbs with the waist, greatly expanding the operating space. For example, this is how it dances the seaweed dance:

▲Robot S1 performs seaweed dance

Such freedom and flexibility are inseparable from the S1's own hardware and software design and the generalization capabilities brought by AI.

The S1 adopts a wheeled chassis + humanoid upper body design, which is more stable and faster than bipedal robots. Its maximum terminal moving speed reaches 10m/s, which is much higher than bipedal humanoid robots. The Zhiyuan Yuanzheng A2 robot, which was just released yesterday, has a walking speed of 1m/s; the Yuanzheng A2-W, which also has a chassis + humanoid upper limb design, has a walking speed of 1.5m/s.

▲Robot S1 shooting

In terms of load-bearing capacity, the S1 single arm has a rated load of 10kg and a self-weight ratio of 1:1.

The S1's hands use a "clip hand" design, which is different from the dexterous hands currently used by many humanoid robots. When Zhidongxi asked whether the robot's appearance design should be more anthropomorphic, what was Stardust Intelligence's plan and what were its considerations?

Lai Jie said that the two-finger gripper can currently complete most tasks and is more practical. Combined with the team's advantage in "force", the gripper can sense and control better. Most of the existing dexterous hands on the market have not yet achieved a balance between flexibility and practicality. The Stardust team is actually developing dexterous hands and will release them in the future. To achieve dexterous operation, one of the key technical points is the tactile sensor. Stardust co-founder Dai Yuan has more than ten years of R&D experience in this field.

2. Take the force feedback route and develop it yourselfAISoftware and hardware integrated system

Stardust Intelligence has independently developed an integrated hardware and software system architecture for AI, deeply integrating AI and robotics technology, combining the advantages of both to jointly create embodied intelligence.

In terms of AI, with the support of a large multimodal model, S1 has the ability to perceive, recognize, and make real-time decisions in complex environments, as well as intelligent understanding and multimodal interactive execution capabilities, and can achieve general operations at the object, task, and environment levels.

Among them, Stardust Intelligence has achieved a key breakthrough in data collection for embodied intelligence.

Data is very important for AI capabilities, but the training data of robot models is not as widely available online as large language models, nor is it available in real time like tens of millions of real cars every day for autonomous driving. In addition to the amount of data, the data dimension is also crucial for robots. How to collect information such as strength is a difficulty faced by the entire industry.

Through a variety of sensors and actuators, S1 can use existing real-world video data and human motion capture data at a low cost, and can also collect high-quality data in multiple dimensions such as touch, force, vision, and hearing from a first-person perspective. Combining these data, S1 can conduct more efficient large-scale training, significantly reducing the cost of high-quality data collection, data volume, and the difficulty of training new tasks, and improving generalization capabilities.

Lai Jie revealed that the company is also expanding the establishment of data factories and may cooperate with other companies to build them in the future.

In terms of robot hardware, S1 adopts a rigid-flexible coupling transmission mechanism design. Through force feedback, sensors monitor the transmission of force in real time, and precisely control the output of control force by sensing the size of force like a human, rather than relying on trajectory estimation. This greatly improves the safety of S1, and it can avoid hurting people, objects, or itself during movement, laying the foundation for the implementation of the scene.

three,16Years of experience in robot R&D, the first batch of products landed in scientific research scenarios

Stardust Intelligence was founded at the end of 2022 and is headquartered in Shenzhen. Its main business is to produce AI robots that can learn, think and work like humans based on self-developed AI technology and robot hardware.

The core founding team of Stardust Intelligence consists of 6 people, all from Tencent RoboticsX Robotics Lab. Founder and CEO Lai Jie has 16 years of experience in robotics research and development. He graduated from Xidian University with a bachelor's degree and later obtained a master's degree in intelligent systems from Wuyi University. From 2012 to 2014, he worked as an assistant researcher at the Hong Kong Polytechnic University, then joined the Baidu Xiaodu Robotics Team, and joined the Tencent Robotics Lab as the No. 1 employee.

In January this year, Stardust Intelligence completed the research and development of the prototype after more than a year, and released the AI ​​robot Astribot S1 for the first time in April, demonstrating its ability to autonomously fold clothes, sort items, vacuum and clean through intelligent human-computer interaction with the support of a large multimodal model.

▲Astribot S1 folds clothes

The new generation Astribot S1 released this time is a product that has been further iterated and optimized on this basis, and will be available for trial to some seed customers.

Why did you choose to start an AI robot business at the end of 2022? What industry judgment did you base it on?

Lai Jie told Zhidongxi that in the more than ten years he has been engaged in the robotics industry, he has always hoped that it can reach the heights it is today. But the reality is that the development of robots has been relatively slow before. The reason why he decided to start a business in 2022 was that he saw the SayCan algorithm proposed by Google in April of that year, which achieved a breakthrough in the upper model, and he had a hunch that a model similar to GPT would appear.

He believes that the upper layer has models that can understand logic, the middle layer has models that can understand the physical world, and the bottom layer has hardware that can interact with the real world. Only in this way can the entire system be truly intelligent.

In the two years since he started his business, Lai Jie has witnessed tremendous changes in the field of AI robots. He said that the development in the past two years has exceeded his expectations, and he is very excited to see such rapid development of the entire industry. The motivation for the entire Stardust Intelligence team is far greater than the pressure.

In his opinion, the combination of large models and robots is a change of the previous step, and the next step is the trend of producing a world model after the combination of the two.

Stardust Intelligence will explore the path to product implementation in multiple scenarios in parallel, including scientific research, B-end, and C-end, among which scientific research is the first step. Lai Jie revealed that the company has reached cooperation with several top universities at home and abroad, and the first batch of products delivered in December will enter university laboratories to build a data flywheel through practice.

Conclusion:AISafety and economy are two major challenges for robots to be put into use

AI robots are increasingly being deployed, such as the humanoid robots of domestic startup Zhiyuan Robotics and American startup Figure, which have entered BYD and BMW factories to “tighten screws” respectively. Stardust Intelligence will also achieve mass production by the end of this year and plans to commercialize it on a large scale next year.

However, safety and economic efficiency are still important challenges faced by AI robots at the implementation level. How to ensure safe interaction between robots and humans and reduce costs to achieve a wider range of implementation are difficult problems that robot manufacturers, upstream and downstream companies, and the entire supply chain need to jointly solve.