news

Zhihui Jun released 5 humanoid robots at once! Developers can also buy them for free

2024-08-18

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Just now, after a year of "pigeon"Zhihuijun, finally withEmbodied IntelligenceNew products are here to fill the gap!

Nothing can better demonstrate confidence in a product than winning one through a raffle.



At the beginning, Zhihuijun did not hide anything. In the "studio" built at the live broadcast site, the robot showed off its operation of mixing drinks under voice commands:





△2x speed

The host of the scene also releasedExpedition A2The robot takes charge.



After a set of small combos, netizens like:



I am the one who is in charge of traffic in the technology section of Bilibili, and I am not kidding you (dog head).



Although the press conference lasted less than an hour, Zhihui and his company Zhiyuan Robotics have been working hard for a year. This time, they really brought a lot of useful information. Let me give you a small summary first:

  • Released the Yuanzheng A2 series of humanoid robots
  • The full-stack open source Lingxi X1 series robot features a "humanoid robot made by humans"
  • The self-developed joint module has been upgraded to mass production, and the degree of freedom of the dexterous hand has been upgraded to 19
  • Defining the evolution path of embodied intelligence from G1 to G5

In short, it can be said that while showing off the product's strength, it also maximizes the hardcore DIY technological sentiment.



Let's talk about the specific details one by one~

Humanoid robots for mass production

Just as Zhihuijun himself said, since the release of the Yuanzheng A1 in mid-August last year, he and the Zhiyuan Robot behind him have been rather low-key, and it can be said that they have been "delayed" from the public eye for a year.



But in fact, during this year, Zhiyuan Robotics has been "quietly doing great things."

Jiang Qingsong, General Manager of Sales and Service of Zhiyuan Robotics, added:

We iterate every month, but because we do not have the conditions for commercial use, we do not promote it.

Now, facingCommercialZhiyuan did give some interim thoughts and answers at this press conference.

First of all, in terms of products, the main feature of the Yuanzheng A2 series robots is a "Improvements for mass production”。



To this end, Zhiyuan has released three robots suitable for different scenarios:

  • Yuanzheng A2, 1.7 meters tall and weighing 140 kilograms, is an interactive service robot that can walk on two feet. With the support of a large model, it has smooth explanation capabilities and stable movement functions.
  • Yuanzheng A2-W is a wheeled flexible intelligent manufacturing robot. It is the one that made drinks for Zhihui at the beginning. It has the ability to dynamically arrange tasks, execute complex operations, and work with two arms in coordination.
  • Yuanzheng A2-Max is a heavy-duty special robot that can move objects weighing more than 40kg. This robot is currently in the product development stage.



In terms of specific technical details, Zhiyuan Robotics divides the robot system into power domain, perception domain, communication domain and control domain.



Power DomainIn terms ofPowerFlowThe joint module has achieved iterative upgrades for mass production, with the strongest model having a rated torque of 270Nm and a peak torque of 512Nm.



Zhihui Jun said that the improvement of PowerFlow is mainly due toreliabilityandstabilityAt the scene, he also revealed that this joint module will be available for sale.

In addition, robotsDexterous HandsThe degrees of freedom have also been upgraded to 19, of which 12 are active degrees of freedom, and tactile perception and visual-tactile perception technologies based on MEMS principles have been introduced.

Use tools to tighten the screws:



Unpacking the courier:



It is no problem to do delicate work of threading a needle:



Perception DomainIn terms of driving, the Yuanzheng A2 series integrates sensors such as RGBD cameras, lidar, and panoramic cameras, introduces an autonomous driving occupancy perception solution, and further improves environmental understanding capabilities through SLAM algorithms.

Communication domainIn terms of embodied intelligence, Zhiyuan has developed a lightweight, high-performance communication framework specifically designed for embodied intelligence.AimRT

Compared with third-party middleware such as ROS, it has improved performance, stability, system deployment efficiency and flexibility, and is compatible with the existing ROS/ROS2 ecosystem.

AimRT will also be open sourced at the end of September.



Control DomainIn terms of the two aspects, by combining the model-based and learning-based algorithms, Zhiyuan has further improved the robot's motion control and adaptability. It has also pre-researched a natural language instruction set driven by different robot bodies.AgentOS, based on reinforcement learning, accurate orchestration and efficient execution of robot skills can be achieved.



In addition, Zhiyuan has also built an open ecosystem-oriented software platform AIMA (AI Machine Architecture), covering on-machine, cloud and client. The platform integrates the core technology of robot software and provides a rich secondary development interface.

Evolution of Embodied Intelligence Technology

It is worth noting that in line with its own commercialization process, Zhiyuan has also proposed similar autonomous driving L1-L5Evolution of Embodied Intelligence Technology

G1 to G5 (where “G” refers to general artificial intelligence).



Among them, G1 isTraditional automation stage, the robot performs tasks based on a large amount of manual choreography, assisted by simple vision,Basically no generalization ability

In the G2 stage, with the support of large models, robots began to realizeRapid migration of a large category of scenarios

The approach is to extract reusable atomic capabilities for a large number of different scenario tasks and implement these capabilities in a relatively general way.

In the past year, Zhiyuan Robotics has made a phased breakthrough in the G2 route, realizing a series of zero-shot and few-shot general atomic skills, including the pose estimation model UniPose, the grasping model UniGrasp, and the force control plug-in model UniPlug.



However, a lot of manual arrangement was still used in the above stages.

arriveG3 End-to-end Phase, robot training fromAlgorithm-drivenbecomeData-driven

At this point, the manual design of various atomic capabilities has changed to relying on large amounts of data collection. The goal of this stage is to form a set ofGeneral Skills Training Framework

For example, the "PPT ability" repeatedly mentioned in the press conference, namely, grab, release, and pass...



Zhihui Jun specifically reminded that current experiments have shown thatEnd-to-end is a route with the potential to scale up

Based on G3, G4 mainly realizesUniversal Operational Model

For example, for the two actions of opening a bottle and opening a door handle, the G3 stage still needs to collect data separately to train different skills.

But for humans, the underlying logic of actions is actually the same.

Therefore, G4 introduces a large amount of real and simulated data across scenarios, combined with similar technologies such as time models, to enable AI to understand the physical principles behind it and perform more complex tasks.

As for the ultimate goal: AGI, for embodied intelligence, the main thing is to be able to go beyond the laboratory and have the ability to generalize across tasks in open scenarios.

From G1 to G5, a notable trend is:Rising demand for data



Jiang Qingsong revealed that Zhiyuan has now entered the G3 stage. On the G3 route, Zhiyuan Robot has also developed a set of embodied data solutions.AIDEA

This is because, “there is as much intelligence as there is artificiality, and this still holds true in the field of embodied intelligence.”

In other words, high-quality human-machine operation data is very critical.



Specifically, AIDEA includes data acquisition body, remote control equipment and data platform.

The data acquisition body is a reliable and stable robot with rich types such as wheeled and foot-type robots.

The teleoperation device supports full-body mapping, arm-hand coordination, and high-precision real-time.

The data platform covers the entire chain from data collection, data labeling, data management, model training, model evaluation, model deployment to data feedback, and supports SaaS services and private deployment.



The easter egg is that based on AIDEA's millions of real machine and tens of millions of simulation data,Open Source. In the fourth quarter of this year, interested friends can squat down~

"Humanoid robots are made by people"

Did you think that after introducing the Expedition series, the annual Zhiyuan Robot Conference is over?

No!

At a command from Zhihuijun, Expedition A2-W slowly pushed up a mysterious black box that was not as tall as it.

It looks familiar, it seems to be the thing that Zhihui commanded A2-Max to move away at the beginning of the live broadcast:



The thing inside broke out of the box, and it was another robot launched by Zhiyuan.

Lingxi X1

Born in the Zhiyuan X-Lab laboratory (which can also be simply understood as the Zhihuijun laboratory within Zhiyuan), Zhiyuan Robotics was specially established for extreme innovation and agile exploration.



The project was launched in June this year, with a team of 10 people working on it. It was officially launched less than half a month ago.



Video footage shows that Lingxi X1 has been through many falls:



Now I can handle being pushed hard without falling.

△Someone come here, someone is bullying the embodied robot (not)

Lingxi X1 is 1.3 meters tall and weighs less than 33kg - compared to the Expedition series, the Lingxi series looks smaller and more exquisite; its maximum walking speed is higher than 2 meters per second, and the single-arm load is greater than 3kg.

It is reported that the entire body of Lingxi X1 is constructed with two self-developed PowerFlow joints, named PF86 and PF52 respectively.

These two PowerFlow models are sold externally.



In addition, Lingxi X1 also has many functions such as hollow wiring, absolute value encoding at the output end, and support for PF-Link smart interface.

It incorporates a modular design concept and can be easily disassembled and assembled through simple clamps, especially its two "hands".

You should know that the standard configuration of Lingxi X1 is self-developed by Zhiyuan X-Lab.Two adaptive universal grippers with feedforward force control and ultra-low cost

Adaptive grip, only 1 active degree of freedom.



But it is highly generalizable and can capture anything.



Even a needle lying flat on the table can be picked up steadily.



But the actuator of Lingxi X1 can not only be an adaptive universal gripper,It can also be a six-dimensional force sensor with a cost of hundreds of yuan, the kind that meets the needs of different scenarios.



What’s interesting is that Lingxi X1 can turn on “machine mode”.

To put it simply, it is to open the brain of Lingxi X1 (which was originally empty), install a mobile phone in it, and use a mobile phone with higher cost performance, richer functions, stronger computing power and wider popularity as its brain.

Then you can see the following funny and a little weird dialogue screen:



In addition, the Lingxi series also has robots:

Professional Data Acquisition Robot X1-W

Like the A2-W in the expedition series, they are all wheeled, and the upper body can be raised and lowered.

However, compared with the A2-W all-in-one robot for data collection, training and reasoning, the X1-W is cheaper and is suitable for only real data collection and serving the pre-training stage. The A2-W can serve multiple processes including data collection, training and reasoning, and is not only useful for pre-training but also for fine tuning.

Different robots can be selected for collection in different application scenarios to meet the full-stack data collection needs.



Finally, the price is something that everyone is concerned about.

"Companies are very competitive, we will not participate in price wars." After saying this, Zhihui Jun pressed the PPT remote control in his hand, and the price of Lingxi X1 appeared on the screen -

0 yuan.



He said that in addition to the self-developed joints, Lingxi X1's main design drawings, software framework, middleware source code, and basic operation and control algorithms will all soon be open source.

The main feature is a "humanoid robot made by everyone".



Zhihui also advertised for his X-Lab on the spot, "Welcome young people with ideas to join Zhiyuan and X-Lab."



One More Thing

Finally, let’s briefly and formally introduce AgiBot.

It was founded in February 2023 and is committed to the integrated innovation of AI+robotics, as well as the creation of embodied intelligent robot products and application ecosystems.

The founding team includes Zhihuijun (Peng Zhihui), who, after graduating from the University of Electronic Science and Technology of China, worked at OPPO Research Institute and Huawei (joined Huawei as a talented young man). At the same time, he is also a B station up master, calling himself Wild Iron Man.



Other founders include Chief Scientist of Smart Healthcare at Shanghai Artificial Intelligence Research Institute and Professor of the School of Mechanical and Power Engineering at Shanghai Jiao Tong University.Yan Weixin

In August last year, Zhiyuan Robotics launched its first product:

A prototype of a general-purpose embodied intelligent robot, A1 of the Expedition series.



Today marks the second time that Zhiyuan Robot has publicly displayed its R&D and product progress.

As of now, Zhiyuan Robot has completed multiple rounds of financing including the A4 round.

Zhiyuan officials also revealed:

This year, Zhiyuan Robotics estimates that its shipments will reach about 300 units, including about 200 bipedal units and about 100 wheeled units.