Beijing Humanoid Robot macht Auftakt auf der WAIC: Das erste industrielle System für autonome kooperative Arbeiten mit mehreren Ontologien, Szenarien und Aufgaben im ganzen Land.
On July 26, the World Artificial Intelligence Conference (WAIC) kicked off in Shanghai. As a global benchmark for artificial intelligence technology, this year's conference has gathered top - tier enterprises and research institutions from over 30 countries around the world to jointly explore the cutting - edge breakthroughs and industrial transformations in artificial intelligence. The Beijing Humanoid Robot Innovation Center (hereinafter referred to as "Beijing Humanoid") made its debut with the nation's first real - world demonstration of "multi - body, multi - scenario, and multi - task autonomous collaborative operations" in the industrial field. This indicates that the "Huisi Kaiwu" system, launched in March this year, has achieved a comprehensive upgrade in the capabilities of its distributed embodied intelligent agent system.
The Beijing Humanoid Robot Innovation Center presented the nation's first real - world demonstration of "multi - body, multi - scenario, and multi - task autonomous collaborative operations" in the industrial field at WAIC.
Currently, embodied intelligence remains in the early stage of "single - scenario, single - task, and single - configuration". The industry is in urgent need of a general and scalable platform to support the application of multi - configuration robots in complex real - world environments. The collaborative demonstration of multiple industrial - scenario embodied intelligent agents brought by Beijing Humanoid this time is the industry's first publicly demonstrated multi - embodied intelligent agent collaborative application using real machines, representing a significant breakthrough. This system is built on the general embodied intelligent platform "Huisi Kaiwu", featuring cross - body compatibility and asynchronous task - collaborative autonomous scheduling, truly realizing "one brain, multiple machines; one brain, multiple functions". This also marks that embodied intelligence is moving from the laboratory to industrialization, providing a practical and replicable path for real - world applications.
Meanwhile, this series of embodied intelligent agent capabilities are also being applied to Beijing Humanoid's star product, "Tianguang", continuously driving its technological evolution and real - world application expansion. At the WAIC exhibition area, "Tianguang" won wide recognition from on - site audiences with its wonderful dance performances. Performing "dynamic dances" is a comprehensive demonstration of a humanoid robot's whole - body motion control ability. For a full - sized humanoid robot to dance, the overall difficulty increases exponentially due to the increase in height and weight, making it extremely challenging to complete complex whole - body coordinated movements smoothly. The smooth and coordinated dance movements of "Tianguang" on - site fully demonstrate its leading technological strength in whole - body coordination and dynamic control. In addition, "Tianguang" is equipped with a dual - battery quick - change power system, enabling long - term battery life. Its comprehensive energy - efficiency indicators are at the international leading level, and its dual - arm end can bear a greater steady - state load. In the future, it can perform various fine - operation tasks in industrial and logistics scenarios. At the 2025 World Robot Conference, which will open on August 8, the newly upgraded "Tianguang" will focus on demonstrating its core technologies, such as autonomous navigation ability with network environment perception and whole - body control - based autonomous operation, accelerating the large - scale implementation of humanoid robots in real - world scenarios.
The star robot "Tianguang" of Beijing Humanoid was surrounded by audiences at the WAIC exhibition area.
General Embodied Intelligent Platform Schedules Multiple Intelligent Agents to Achieve the Nation's First Fully Autonomous Multi - Body Collaborative Task Chain
Beijing Humanoid presented a complete industrial - scenario task chain at the WAIC exhibition area: Under the unified command of the general embodied intelligent platform "Huisi Kaiwu", each intelligent agent robot autonomously reports progress and communicates with others. The electrical control cabinet operation area simulates delicate operations, the light - bulb quality inspection line conducts industrial inspections, and the item packaging area performs packaging operations. Robots shuttle between the three scenarios to complete various handling tasks. Through a distributed embodied intelligent agent architecture, the system deploys four task - executing intelligent agents, namely the "Electrician Master", "Porter", "Quality Inspector", and "Packaging Assistant", on four robot bodies. Meanwhile, a task - scheduling intelligent agent is deployed in the cloud to receive instructions and schedule overall tasks, distributing subtasks to the four robots for collaborative completion. Multiple intelligent agents on - site jointly construct an efficient "task - execution network". Operators only need to send a "Start" instruction on the "Huisi Kaiwu" App, and audiences can witness the entire process from task distribution to the collaborative operation of multiple robots on the large screen, directly experiencing the implementation of intelligent robots in various complex scenarios.
The "Huisi Kaiwu" App demonstrated in real - time on - site, allowing audiences to directly see the entire process from task distribution to the collaborative operation of multiple robots on the large screen.
In a common intelligent agent workflow, only one or some intelligent agents are activated at the same time, only supporting synchronous workflows. However, the "Huisi Kaiwu" system enables all intelligent agents to be in an active and asynchronous collaborative state: Robots autonomously execute tasks based on task priorities and their own states, and can communicate with each other using natural language rather than program code, just like humans. After completing a task, they trigger the next - stage collaborative process, truly realizing "machine - decision - making by machines". Asynchronous collaboration of multiple embodied intelligent agents requires the system to have real - time multi - task concurrent processing capabilities, including multi - task insertion, suspension, and resumption with priorities. Compared with synchronous workflows, it poses a higher - level challenge to the overall system. Each intelligent agent robot is equipped with an independent "brain" and "cerebellum", enabling them to work independently and collaboratively, with the full - chain ability from cognition to execution, building a core technological foundation to support future multi - scenario generalization and cross - robot collaboration.
That's why audiences were able to closely observe the smooth and seamless industrial operation processes of robots of various configurations at the exhibition area. Behind this is the continuous evolution and comprehensive display of the R & D team of Beijing Humanoid in key technologies such as distributed embodied intelligent agent systems, multi - modal operation VLA large models, visual understanding, spatial perception, complex task understanding, precise task decomposition, tool invocation, dual - arm collaboration, autonomous error correction, and real - time replanning and execution.
In the electrical control cabinet operation area, after receiving instructions from "Huisi Kaiwu", the robot autonomously completes multi - process delicate operations such as opening the cabinet, pressing buttons, and closing switches. Behind these seemingly simple operations lies the technological achievement of a 35 - degree - of - freedom robot system and a VLA (Visual - Language - Action) model for dual - arm dexterous hand collaboration in long - term, contact - intensive tasks. In the power maintenance scenario, the robot needs to address complex challenges such as long - term task error accumulation, electrical operation logic, and contact with dense components.
The robot autonomously completes multi - process delicate operations such as opening the cabinet, pressing buttons, and closing switches in the electrical control cabinet operation area.
To overcome this challenge, the R & D team developed a dual - arm - hand dexterous operation algorithm based on a visual - tactile - action model by collecting high - quality manual demonstration data through a multi - modal perception system. To address the current challenge of robots' inability to accurately complete long - range tasks, a multi - task serial and prediction algorithm based on key frames was adopted to effectively reduce error accumulation. Through a task self - error - correction mechanism, the model can automatically determine whether an operation is completed based on electrical signals. Combined with fingertip tactile feature extraction, the model can accurately sense contact changes, effectively coping with sudden stress changes in complex components such as knobs and air switches. Compared with mainstream single - modal perception systems and traditional imitation learning, this method can significantly increase the task success rate to over 90%. With these technologies, the robot can adapt to multiple scenarios and make intelligent decisions, ensuring operation safety while promoting the intelligent upgrade of the power industry and reducing high - risk manual operations.
After the electrical control cabinet is powered on, the dual - machine collaborative process initiated in the light - bulb quality inspection area represents a more challenging and flexible display. The robot grabs the light - bulb box from the shelf and accurately transports it to the designated position. Meanwhile, the robotic arm is awakened and starts the standardized operations of "grabbing - inserting - lighting - quality inspection - sorting". The visual sensor, combined with a multi - modal large model, conducts light - effect inspections to determine the quality of each light - bulb and plan the next operation.
The robots in the light - bulb quality inspection area can autonomously judge and sort each light - bulb based on the detected light effects and subtle features.
Throughout the process, based on three - dimensional environmental perception, the robot enables the two robotic arms to autonomously avoid obstacles and dynamically plan grabbing and operation trajectories. Combined with force - position hybrid control technology, the end - effector can sensitively adjust force and position to ensure a smooth insertion process without damaging the light - bulbs. A high - precision visual positioning system simultaneously ensures accurate positioning of the light - bulbs. In addition, through the VLM scenario understanding ability, the robot can autonomously judge and sort each light - bulb based on the detected light effects and subtle features. Under the coordination of the "Huisi Kaiwu" general embodied intelligent platform, the system can autonomously plan processes, flexibly invoke sensors and tools, and achieve multi - device collaboration and rapid task reorganization, providing an efficient and scalable solution for a "flexible automated factory".
In the item packaging area, based on Beijing Humanoid's self - developed end - to - end VLA model, industry - leading real - time error - correction and smooth dual - arm collaborative operations are achieved. The robotic arm performs item packaging and QR - code pasting tasks. Complex disturbances such as lighting changes and item position changes are introduced on - site to simulate a dynamic environment in a real - world scenario. The system can still adaptively handle these interferences and stably complete various operations, truly achieving "mission accomplished". Finally, the robot shuttles autonomously in the exhibition area to complete various item - handling tasks.
This is due to the self - developed VLA model integrated into the central control decision - making module of "Huisi Kaiwu", which endows the robot with the ability to quickly predict and issue actions, ensuring smooth operations at each stage. Under external light interference, robots are prone to action errors or abnormalities. To address this, the R & D team used feature extraction capabilities and data generation solutions to effectively solve the lighting interference problem that often occurs in real - world scenarios. At the same time, asynchronous reasoning significantly improved the reasoning speed, enabling the entire model to have excellent environmental adaptability and execution robustness. When facing sudden interferences, it can real - time sense the task status, autonomously adjust strategies and correct errors, ensuring stable and accurate operations. Taking the QR - code pasting process as an example, the model can understand the current posture of the item bag like a human, quickly predict and issue appropriate actions, enabling the robotic arm to turn and adjust and accurately paste the QR - code in place. Compared with the operating mode of traditional robots that rely on a stable environment, this architecture has demonstrated stronger task - completion ability and environmental adaptability under dynamic conditions through verification in unstructured scenarios.
In the item packaging area, smooth dual - arm collaborative operations accurately complete item packaging and QR - code pasting tasks.
Whole - Body Motion Control and Data Empowerment: Building a New Foundation for Humanoid Intelligence
On the other side of the exhibition area, the humanoid robot "Tianguang 2.0" was dancing vigorously to the dynamic music, instantly igniting the on - site atmosphere and attracting a large number of on - lookers. During the dance, its more than 42 joints with degrees of freedom coordinated perfectly, with the dexterous fingers, fast and smooth arms, and stable swaying torso all demonstrating the advanced and stable whole - body high - dynamic coordination control ability.
Tianguang 2.0 is about 173 cm tall and weighs about 73 kg. The larger weight and longer limbs mean greater motion inertia and torque changes, requiring faster and more accurate real - time calculations and stable responses for motion control. The dynamic linkage of 42 high - degree - of - freedom joints tests the system's limit in multi - joint collaboration, real - time planning, and anti - interference ability. During the high - speed dance, Tianguang 2.0 achieved continuous whole - body center - of - gravity stability and high - dynamic balance control through its large - range lower - limb motion, precise upper - limb control, and flexible torso posture. In addition to its excellent performance in motion control and system collaboration, its longer battery life and more flexible upper - limb operations also provide a solid foundation for its future wide - scale application in industrial and logistics scenarios.
Tianguang was dancing vigorously at the exhibition area, with its smooth dance moves igniting the on - site atmosphere.
It is worth mentioning that the champion robot "Tianguang Ultra", which set a world record at the Beijing Yizhuang Half Marathon in April this year with a time of 2 hours, 40 minutes, and 42 seconds, also made a surprise appearance, attracting a large number of audiences to take photos with it.
The data foundation supporting the operation of this system is also remarkable. At the on - site data collection area, audiences could experience the real - time demonstration of Beijing Humanoid's self - developed isomorphic - arm remote - controlled real machine and simulated Tianguang. Every movement of the operator could be accurately reproduced and completely recorded by the real machine and the simulated robot. This high - dynamic and high - fidelity data collection provides valuable materials for algorithm training. Combined with Beijing Humanoid's self - developed data collection platform, it fully demonstrates the full - chain technological ability from data acquisition through virtual - real combination to model training. These data not only support the rapid iteration of algorithms but also lay a solid foundation for the standardized development of the industry. The RoboMIND dataset released by Beijing Humanoid has been downloaded and used over 40,000 times, becoming an important open - source resource for embodied intelligence research.
