news

Zhiyuan released 5 humanoid robots and proposed the evolution route of embodied intelligence technology from G1 to G5

2024-08-18

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Tencent Technology News: On August 18, Zhiyuan Robotics held the "Zhiyuan Expedition Commercial Launch" 2024 New Product Launch Conference. Zhiyuan co-founder Peng Zhihui hosted and released the "expedition"and"Telepathy"A total of five new commercial humanoid robots in the series——Yuanzheng A2, Yuanzheng A2-W, Yuanzheng A2-Max, Lingxi X1andLingxi X1-W, and demonstrated thePower, perception, communication, controlResearch and development achievements in four areas, as well as embodied intelligenceG1 to G5 Technology RoadmapandAIDEAEmbodied Intelligent Data Systems.

At the conference, Zhiyuan Robotics announced its open source plan, including: High-performance communication framework for intelligent robotsAimRTWill be open source at the end of September, incubated by X-LabLingxi X1The full stack will be open sourced in September, and the AIDEA-basedMillions of real machine and tens of millions of simulation data setsIt will be open sourced in the fourth quarter of this year.

Among them, the Yuanzheng A2 series robots are about 1.7 meters tall. Jiang Qingsong, partner and vice president of marketing at Zhiyuan Robotics, said, "This type of humanoid robot is priced at around 600,000 to 700,000 yuan on the market, and costs around 200,000 to 300,000 yuan." However, Zhiyuan decided not to engage in price wars on the Lingxi series robots, which are about 1.3 meters tall, and opened up the design drawings of the full-stack open source robots.

Zhiyuan Robotics estimates that its shipment volume in 2024 will reach about 300 units, including about 200 bipedal units and about 100 wheeled units, all of which will be shipped in October.

This time, Zhiyuan released 5 robots

At last year's press conference, Zhiyuan released the "Yuanzheng A1" humanoid robot. This year, Peng Zhihui announced two series, "Yuanzheng" and "Lingxi", with a total of 5 robots, which have both wheeled and footed forms, covering application scenarios such as interactive services, flexible intelligent manufacturing, special operations, scientific research and education, and data collection.

Specifically, the "Expedition" series has three robots, namely, the Expedition A2, the Expedition A2-W and the Expedition A2-Max; the "Lingxi" series has two robots, namely, the Lingxi X1 and the Lingxi X1-W. Among them, the Expedition A2-W is a wheeled robot. Jiang Qingsong believes that bipedal robots are the future, but "wheeled + robotic arm" is currently the most widely used type of robot in industrial application scenarios.

Specifically:

• Yuanzheng A2: An interactive service robot with anthropomorphic configuration and human-factor design, multi-mode interaction and intelligent experience, and can move and walk autonomously. At the press conference, Yuanzheng A2 appeared as the host and introduced himself.

•Expedition A2-W: A flexible intelligent robot with dual-arm collaboration capabilities. In the press conference situational drama, A2-W demonstrated operational capabilities such as dynamic task scheduling, complex operation execution, and dual-arm collaborative operations.

•Expedition A2-Max: A heavy-load special robot. At the launch event, it was shown to be able to move a 40kg aviation box. It is currently in the product development stage.

• Lingxi X1: A full-stack open source robot with a series-parallel hybrid configuration arm and differential drive double shoulder joint design, suitable for human-computer interaction and light service scenarios, and fully open source.

• Lingxi X1-W: A professional data acquisition robot that can reduce data acquisition costs and conduct large-scale data collection.

The key capabilities of humanoid robots are reflected in two aspects: "brain" and "movement" operations. The "brain" can help robots understand the world and arrange tasks, while the "movement" involves the robot's operational capabilities in actual tasks and involves the coordination of software and hardware.

At last year's press conference, Peng Zhihui demonstrated that the brain and software of "Expedition A1" used the embodied intelligent brain EI-Brain, the unified software framework AgiROS and the language task model WorkGPT, and independently developed core components and technologies in hardware, such as the core joint motor PowerFlow and the dexterous hand SkillHand.

This year, Peng Zhihui demonstrated the upgrade of the software and hardware of the "Expedition" series at the press conference, but divided the core joint motor, dexterous hand, and software system into "four domains" to describe the updates and changes compared with the past, namely: power domain, perception domain, communication domain, and control domain:

In the power domain module, the intelligent elementPowerFlow joint moduleAfter iterative upgrades and mass production, compared with last year's PowerFlow peak torque of 350Nm, this year's 15022 series joint peak torque is as high as 512Nm.

Dexterous hand degrees of freedomIt has increased to 19, while the Expedition A1 robot released a year ago has 17 degrees of freedom, including 12 active degrees of freedom and 5 passive degrees of freedom; this year's robot dexterous hand has 19 degrees of freedom, of which the active degrees of freedom have doubled to 12, and tactile perception and visual tactile perception technology based on MEMS principles have been introduced; high-precision force-controlled 7-degree-of-freedom dual arms can accurately perform impedance control, admittance control, force-position mixed control and other force control tasks, and also support dual-arm drag teaching mode and visual adjustment.

In the perception domain, sensors such as RGBD cameras, lidar, and panoramic cameras are integrated to introduceAutonomous drivingOccupancy perception solution further improves environmental understanding capabilities through SLAM algorithm.

In the communication domain, we developed our own embodied nativeCommunication framework AimRTCompared with the third-party middleware such as ROS used last year, it has improved 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 the control domain, we combined the Model-based and Learning-based algorithms to improve the robot's motion control and adaptability; we also conducted preliminary research on AgentOS, which is driven by natural language instruction sets and can adapt to different robot bodies.

Zhiyuan Robotics has built an open ecosystem-oriented software platform AIMA (AI Machine Architecture), covering on-machine, cloud and client.

Imitating autonomous driving, Zhiyuan proposed the "Evolutionary Route of Embodied Intelligence Technology"

Similar to the evolution levels of autonomous driving "L1-L5", Zhiyuan Robotics also attempts to define the evolution path of embodied intelligent technology from "G1 to G5".

Peng Zhihui explained at the press conference that G1 is tailored for specific scenarios, but cannot perform low-cost and fast migration to different scenarios; G2 has insights into a large number of scenario tasks, and can achieve rapid migration to a certain extent, and can be orchestrated in conjunction with a large language model framework. Robots at this level have certain generalization capabilities; G3 takes an end-to-end approach, and the overall architecture of G3 and G2 has similar algorithmic capabilities, but the driving method has changed from algorithm-driven to data-driven; G4 implements a large general operation model, introduces simulation data and world models, and further improves the robot's performance in complex tasks; G5 level is a long-term traction goal, which can achieve end-to-end operations from perception to execution.

Peng Zhihui said, "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 has been commercially applied in multiple actual scenarios for flexible intelligent manufacturing and interactive service scenarios."

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

"The data acquisition body provides multiple types of reliable and stable robots, such as wheeled and foot-type robots, equipped with a set of remote control equipment that supports full-body mapping, arm-hand coordination, and high-precision real-time. It also includes a full-link data platform such as 'data collection-data labeling-data management-model training-model evaluation-model deployment-data return', supporting SaaS services and private deployment. One million real machine and tens of millions of simulation data sets based on AIDEA will be open source in the fourth quarter of this year."

At the end of the press conference, Peng Zhihui announced that Zhiyuan Robotics had established X-Lab (Peng Zhihui Laboratory), which incubated two new products: the full-stack open source robot Lingxi X1 and the digital acquisition robot X1-W.