news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Once a year, Zhihui’s second generation humanoid robot is here.

Dexterous little hands are making waves at the mahjong table:

No, no, let's take a look at the overall temperament first:

Then there are the routine housework, which can be seen to be very skilled:

In 2024, when big models are all the rage, if we were to ask what is the hottest topic in the field of AI, "embodied intelligence" would definitely be one of them.

Looking at China, Zhiyuan Robotics, a self-sustaining intelligent company, has attracted much attention. Since its establishment in February 2023, Zhiyuan Robotics has completed 6 rounds of financing and quickly became a "top project" in the robot startup circle.

After only six months of starting the business, Zhihui and his team came up with their first product, the "Yuanzheng A1". As soon as it was launched, the walking ability and human-computer interaction of the "Yuanzheng A1" were at the leading level in the industry. However, after the debut of the "Yuanzheng A1", we seem to have never heard of its next step, and even no live video has been released.

But after today's press conference, we know that Zhiyuan Robotics is "doing great things quietly."

At this press conference, Zhihui Jun launched three Yuanzheng series robot products at once: interactive service robot "Yuanzheng A2", flexible intelligent manufacturing robot "Yuanzheng A2-W", and heavy-duty special robot "Yuanzheng A2-Max". In the last "One more thing" session, the modular robot series products "Lingxi X1" and "Lingxi X1-W" incubated by Zhiyuan X-Lab were also officially unveiled as "easter eggs".

Compared with the previous generation, the five robots adopt a family design language, and have serialized the appearance of the robots, combining wheeled and foot-based forms, covering application scenarios such as interactive services, flexible intelligent manufacturing, special operations, scientific research and education, and data collection.

Regarding the progress of mass production, Zhiyuan Robotics also revealed the latest news: the estimated shipment volume in 2024 will reach about 300 units, including about 200 bipedal units and about 100 wheeled units.

New generation expedition A2

Since the launch of the Yuanzheng A1, Zhiyuan’s direction for the implementation of its robot products has been very clear: “working in factories.” This year’s Yuanzheng A2 is even more fully prepared for mass “working in factories.”

Let's take a look at the members of this family:

"Yuanzheng A2" is an interactive service robot with more than 40 active degrees of freedom joints and human-like dexterous hands, which can simulate rich and complex human work. It has a brain that can continuously learn, supported by a large language model. In addition, it also has a multimodal perception and input system, which can perceive the emotions of the interactor through vision.

"Expedition A2-Max" is a heavy-duty special robot that has both great strength and the advantage of dexterity. In the opening scene, it easily moved a 40kg aviation box. Currently, "Expedition A2-Max" is still in the product development stage.

"Yuanzheng A2-W" is a flexible intelligent robot whose two arms can operate independently or work together. It uses a wheeled chassis, combined with the RGBD laser radar, panoramic camera, sensors configured for global safety and other components on the fuselage, allowing it to move quickly and smoothly in various environments.

Before the press conference, the performance of "Yuanzheng A2-W" was like "Jarvis" from Iron Man came to life. After understanding Zhihui's instructions, it opened the Coke bottle completely autonomously, put grapes in the juicer, pressed the juicer, and poured the drink in the juicer into the cup without spilling a drop, serving Zhihui a glass of "grape carbonated Coke".

In addition to the improved appearance, the "Yuanzheng A2" series of robots are also improved inside and out. For the core components of the robot, "Zhiyuan Robot" innovatively divides the robot system into power domain, perception domain, communication domain, and control domain. Obviously, "Zhiyuan Robot" is aiming for a full-stack ecosystem.

In the power domain, Zhiyuan Robotics has achieved mass production and iterative upgrades for the PowerFlow joint module. From the perspective of parameters, the PowerFlow joint module has been greatly improved. In terms of stability and reliability, Zhiyuan Robotics has also conducted a lot of tests and optimizations on its peak performance and aging speed.

As for the dexterous hands that are important for "working", this time the "Zhiyuan Robot" has also been upgraded: the number of degrees of freedom has jumped to 19, the active degrees of freedom have doubled to 12, and more modes have been introduced in terms of perception capabilities, such as tactile perception and visual-tactile perception technology based on MEMS principles.

In addition, "Zhiyuan Robot" has also launched a high-precision force-controlled 7-degree-of-freedom dual-arm, which can accurately perform various force control tasks such as impedance control, admittance control, force-position mixed control, etc., and also supports dual-arm drag teaching mode and visual adjustment.

In terms of the perception domain, the new generation of expedition robot series integrates sensors such as RGBD cameras, lidar, panoramic cameras, introduces cutting-edge perception solutions for autonomous driving occupancy, and further improves environmental understanding capabilities through SLAM algorithms.

In the communication domain, Zhiyuan Robotics has developed AimRT, a native, lightweight, high-performance intelligent robot communication framework. Compared with third-party middleware such as ROS, it improves performance, stability, efficiency and flexibility of system deployment, while being fully compatible with the existing ROS/ROS2 ecosystem. AimRT will be open source at the end of September.

In terms of the control domain, Zhiyuan Robot combines the model-based and learning-based algorithms to further improve the robot's motion control and adaptability. For the model-based, Zhiyuan Robot further improves the robustness of the system, so the robot's demonstration at the press conference was so smooth and silky. For the learning-based algorithm, Zhiyuan Robot has highlighted a key point here, hoping to promote the transformation of robot training methods from algorithm-driven to data-driven.

Based on the operation and control algorithm, "Zhiyuan Robot" has pre-developed AgentOS, which is driven by a natural language instruction set and can be adapted to different robot bodies. Based on reinforcement learning, it can achieve precise orchestration and efficient execution of robot skills.

First proposed G1-G5 embodied intelligence technology evolution route

At this conference, Zhiyuan Robotics also proposed for the first time the technical evolution route in the field of embodied intelligence, which includes five stages from G1 to G5:

G1 refers to the basic automation stage, which is based on manually designed features and simple machine vision to provide some feedback. The deployment of this stage is tailored for specific scenarios. Although it can solve the task execution problems in some scenarios, it cannot be quickly migrated in different scenarios at low cost.

G2 is the "universal atomic skills" stage, which extracts some reusable atomic capabilities for a large number of different scenario tasks and various operation task requirements. In short, it is to achieve rapid migration of similar scenario tasks in a relatively general way, and then coordinate tasks with the large language model framework, so that the robot has a certain generalization ability.

G3 means that the overall architecture of embodied intelligence begins to adjust to an "end-to-end" route. Although the algorithm architecture of this stage may be similar to that of the G2 stage, the way each "atomic capability" is formed at this time has changed from being driven by manually designed algorithms to being driven by data after a large amount of data has been collected.

Different from the first three stages, a universal end-to-end operation model will appear in the G4 stage. At this stage, a large amount of real data and simulation data across scenarios can be introduced, and the world model can be introduced to help AI understand the physical principles behind the real world and the commonalities between the underlying logic behind different tasks. For example, for the two things "unscrewing the bottle" and "unscrewing the door handle", it is no longer necessary to collect two separate data to obtain two separate skills. Eventually, it will lead to the G5 stage that is infinitely close to AGI.

In the past year, Zhiyuan Robotics has made a breakthrough in the G2 route, realizing a series of zero-shot and few-shot general atomic skills, including the universal pose estimation model UniPose, the universal grasping model UniGrasp, and the universal force control plug-in model UniPlug. The atomic capability model in the G2 stage, oriented to flexible intelligent manufacturing and interactive service scenarios, has been commercially applied in multiple actual scenarios.

On the G3 route, Agibot has also developed a complete full-process embodied data solution AIDEA (Agibot Integrated Data-system for Embodied AI).

But we all know that in order to form such a plan, we first need to invest a lot of data collection costs, which not every robotics team has the conditions to do.

Zhiyuan Robotics has achieved this and is preparing to open source it in the fourth quarter of this year.

Zhihuijun said that data collection is the pain point of the entire process. In the past year, they have done a lot of infrastructure work for data and formed a complete, full-process data collection and application solution AIDEA.

In addition to providing hardware solutions, AIDEA also provides the entire cloud full-link data platform, the AIDEA ML machine learning platform for model training and algorithm evaluation, and the AIDEA Sim simulation platform.

Zhihuijun said that he expects that Zhiyuan will have more than 100 freely deployable robots dedicated to end-to-end data collection. They will also open source one million real machine and tens of millions simulation data sets based on AIDEA in the fourth quarter to actively build an open ecosystem.

One more thing by Zhihuijun

Perhaps because he felt that holding a press conference only once a year was too late, Zhihuijun also announced one more thing at the end of the press conference: two new products, the full-stack open source robot Lingxi X1 and the professional data acquisition robot Lingxi X1-W.

These two robots come from X-Lab (Zhihuijun Laboratory), and were made by 10 people in less than 3 months. They have many innovative details. For example, they developed two new PowerFlow joints, PF86 and PF52, which incorporate modular design concepts and can be easily disassembled and assembled with simple clamps.

Based on X-Lab's pursuit of ultimate innovation, Zhihuijun said that the Lingxi X1's main design drawings, software framework, middleware source code, and basic operation and control algorithms will be open source.

In addition, in terms of actuators, X-Lab has developed its own adaptive universal gripper with feedforward force control and ultra-low cost, and has also launched a six-dimensional force sensor with a cost of hundreds of yuan to meet scene requirements.

By the way, Lingxi is priced at zero yuan, and most of its design information and codes are open source, but it only sells parts.

Zhihuijun expressed the hope that this would promote the arrival of an era where "humanoid robots are artificial by everyone."

How do you like it?