Author | Du Lin Editor | Mo Ying
For a bipedal humanoid robot capable of continuously playing badminton with humans autonomously, how much of its original capabilities can be reused when assigned to new tasks?
Zhidx reported on October 18 that at the 140th Canton Fair this year, PHYBOT Technology arranged human-robot badminton and table tennis matches in multiple locations and provided bilingual Chinese-English guided tours at the Canton Fair History Museum. Meanwhile, at its corporate booth, PHYBOT Technology showcased application solutions for scenarios such as heavy-duty transportation, special inspection, and dexterous manipulation.
Prior to this, at the 139th Canton Fair in April this year, this Guangzhou-based company focused on badminton interaction and participated in the first five-day phase. For the 140th Canton Fair, PHYBOT Technology plans to sustain its guided tour services and sports interactions across all three phases, running until November 4.
With increased tasks and extended service hours, the requirements for humanoid robots have become more numerous and stringent: it must solve not only the technical issues in a single match. Whether models can be reused in new tasks and how equipment can be adapted to new venues will jointly determine whether continuous service can be realized.
01.Behind Continuous Rallies, How Do Humanoid Robots Cope with Changes?
Taking the badminton rally of the bipedal humanoid robot PHYBOT C at the 139th Canton Fair as an example, every time the robot returns a shuttlecock, it must complete target tracking, landing point prediction, shot selection, and body coordination.
Different opponents have different movement habits, and the speed, direction, and landing points of the shuttlecock also vary. The robot cannot simply repeat a fixed set of movements, and these unexpected trajectories cannot be preset; it can only rely on the robot to "adapt to changes on the fly."
Especially in the scenario of a bipedal humanoid robot playing badminton, movement, racket swinging, and balance control must be completed synchronously. The robot needs to coordinate its arms and torso while adjusting its stance to ensure body stability during the stroke.
At the same time, while the robot is making judgments, the shuttlecock does not stop. It needs to adjust its movements based on continuously updated position information to complete the stroke at the right position and timing.
Therefore, whether perception, decision-making, and execution can coordinate in a timely manner directly affects the sustainability of the rally.
Sports scenarios thus provide a testing environment for robotics research with clear conditions and sufficiently rich variations: the court and rules have boundaries, but human movements are hard to preset.
Whether capabilities such as target tracking and motion control formed during badminton matches can be reused when switching to another sport is a question developers need to answer. This issue is not limited to badminton; whether it is table tennis, football, or other sports, or even tasks like storage, transportation, and inspection, it is equally necessary to determine whether existing capabilities can be transferred to meet new requirements.
Typically, to achieve cross-task and cross-embodiment reuse, models often require fine-tuning for the target environment. Even if the same task is performed, changing the embodiment form may still necessitate retraining.
In simple terms, this is "one robot, one training." The core pain point lies in the strong coupling of the model with the robot's hardware platform, control interfaces, and task environment, resulting in high strategy transfer costs.
Taking grasping tasks as an example, when switching from a single-arm robot to a dual-arm robot, the number of joints and degrees of freedom configuration, the number and layout of cameras, end-effectors, and motion control methods may all change, making it difficult to directly transfer the original strategy.
Even when using pretrained models, facing embodiment configurations not covered in the training distribution, developers may still need to collect or supplement operational data for that configuration, adjust perception input and motion output interfaces and related modules, and conduct training or fine-tuning to complete embodiment adaptation.
This means that pretrained capabilities can provide an initialization or transfer starting point, but it is not zero-cost transfer. Expanding to each new robot embodiment and task still requires new data collection, embodiment adaptation, and validation work, thereby increasing development investment and extending deployment cycles.
Regarding the transfer of such robotic capabilities, the industry has been conducting continuous research.
As early as 2023, Google DeepMind, in collaboration with 33 academic laboratories, launched the Open X-Embodiment project. It integrated over 1 million real-world manipulation data points from 22 robot embodiments, covering 527 skills and over 160,000 tasks, and used this data to train the RT-X series models, verifying whether the operational experience accumulated by one robot could help other robots complete tasks. [1]
PHYBOT Technology is also exploring the reuse of existing capabilities. According to public information, PHYBOT Technology took less than a month to transfer the capability of its humanoid robot playing badminton with humans to the table tennis project.
This case demonstrates that there is room for reusing capabilities such as target tracking and motion coordination in both types of projects, and it also addresses the industry's pain point of "one robot, one training," namely how to achieve efficient and low-cost capability reuse.
And when applications further shift to specific scenarios such as transportation and inspection, operational condition adaptation will also become a new challenge.
Transportation tasks must deal with changes in payload and end-tools, while inspection tasks must cope with changes in terrain and environment. At the same time, complex operations require solving not only how the body moves, but also the robot must understand work objectives, arrange multiple steps, and adjust plans based on execution results.
It can be said that robots now possess the capability to complete new tasks through transfer; however, currently, single tasks are more common. Next, the industry needs to focus on the robot's understanding, planning, and execution of complex, long-horizon tasks.
02.Large Models Handle Deep Thinking, How to Accelerate Robot Execution and Deployment?
Understanding human instructions and judging relationships between objects involve both language and visual capabilities, as well as the need to understand human intentions, apply physical common sense, and perform spatial reasoning.
The progress of general large models themselves provides new possibilities for robots to understand intentions, achieve deep thinking, and perform task planning.
Gemini Robotics-ER, launched by Google DeepMind in 2025, focuses on spatial understanding and embodied reasoning, capable of planning multi-step tasks and coordinating with existing low-level controllers. The model is responsible for understanding and planning, while the controller handles execution; the two complete robotic tasks through division of labor. [2]
Meanwhile, research on directly enabling general large models to assist robots in completing tasks is also advancing.
In the RoboDojo evaluation released in September 2026, GPT-6 Astra achieved an average success rate of 22.48% across 42 simulation tasks and 2,100 tests. The research also pointed out that it still has shortcomings in tasks such as precise manipulation and dynamic control. [3]
This result demonstrates the potential of general models participating in robotic tasks. Robot companies thus have the opportunity to leverage general large models to provide understanding and planning guidance at the front end, while enabling the existing edge perception, real-time decision-making, motion control, and execution feedback of embodied intelligence to execute collaboratively at the back end, ultimately forming a closed loop in the physical world.
In this regard, PHYBOT Technology holds a similar judgment. Its recently proposed Embodied PhyStack advocates that robot companies do not need to build a centralized, omnipotent "brain" themselves, but rather enable general large models to collaborate with physical agents, motion control systems, and embodiments through division of labor.
Relevant experiments published by PHYBOT Technology further investigate how general large models and robot control systems can coordinate.
The first set of experiments verifies whether a general large model can direct a robot to complete multi-step tasks. In this set of experiments, the model is responsible for understanding goals and planning steps, and then PHYBOT Technology's self-developed motion control system PhyCore controls the specific movements of the robot's body.
In a workpiece grasping task, the robot made two consecutive mistakes while grasping a silver workpiece. After feeding the execution status back to the model and replanning, it completed the grasping task on the third attempt.
The model can analyze mistakes and adjust plans, but this set of tasks took about 3 hours, making it difficult to meet the timeliness requirements of continuous operation scenarios.
To address this issue, PHYBOT Technology's Embodied PhyStack implements a finer division of labor. It subdivides the common "brain-cerebellum-embodiment" into four layers: "General Large Model—PhyAgents—PhyCore—PHYBOT," which respectively undertake deep thinking, real-time decision-making, motion control, and embodiment execution.
The general large model reasons based on task intentions and environmental information, decomposing long-horizon tasks into sub-task sequences; the physical agent PhyAgents makes real-time decisions based on the environment, embodiment state, and task progress through the edge vision-language-action (VLA) model.
PhyCore translates decisions into motion control commands, which are executed by the PHYBOT embodiment; the execution results are then fed back to the decision-making and planning layers to adjust subsequent actions and task arrangements. Thus, the robot achieves a physical closed loop of "perception—decision—control—execution—feedback."
Building on this, PHYBOT Technology conducted a second set of real-robot experiments. This set of experiments recreated a hotel cleaning supplies storage scenario. The robot completed the storage operations according to the sub-task sequence provided by the general large model, with the real-time decision-making support of PhyAgents.
In this set of experiments, the robot took less than 2 minutes to grasp three workpieces. Calculated based on the average time per workpiece, the execution efficiency of the second set of experiments is over 100 times that of the first set.
The purpose of this division of labor is to enable the edge system to respond to on-site changes in a timely manner, reducing the wait for action execution on the step-by-step reasoning of upper-level models.
The work involving general large models can also be extended to the training phase.
In this regard, PHYBOT Technology also conducted a third set of experiments. Engineers input task objectives, constraints, and precautions for sim-to-real transfer. The model generates training agents, calls GPUs to conduct training, and automatically verifies and iterates in simulation before deploying to the real robot once standards are met.
This time, it is the development efficiency that is improved.
The test results provided by the PHYBOT Technology team show that the work efficiency of training for new scenarios has increased by dozens of times, and some work that originally relied on engineers for manual debugging and repeated verification has begun to be completed with the participation of models.
The three sets of experiments involve task planning, real-time execution, and new task training, respectively. They point to a new form of division of labor: general models provide understanding and reasoning capabilities, while robot teams focus on processing feedback, real-time decision-making, and motion control in the physical world, enabling planning to be executed on specific devices.
Accelerating training and deployment is only the first step in putting robots into use. Whether these technological advances can be transformed into continuous services depends on the operational performance of robots in actual scenarios.
03.Multi-Scenario Applications, How to Turn Them into Sustainable Business?
Focusing on actual scenario operations, this summer, we saw the continuous service capabilities of robot companies in shopping malls. They brought humanoid robots, which were previously only displayed at exhibitions and TV programs, to the public, transitioning robots from a one-time performance to continuous service.
PHYBOT Technology is a prime example. It has introduced sports interaction into shopping malls, attempting to transform the rally demonstrations at exhibitions into normalized operational services.
At the 139th Canton Fair, the world's first bipedal humanoid robot capable of autonomously playing badminton with humans made its debut. In early July, PHYBOT Technology brought this robot to a shopping mall in Guangzhou, launching the "GouC Badminton Court" pop-up store for nearly a month.
During the operation, the robot faced real customers of diverse ages, regions, and cultural backgrounds, ranging from the elderly to children, and from Chinese consumers to overseas tourists.
Operational data released by PHYBOT Technology shows that the robot at the "GouC Badminton Court" received up to over a thousand customers in a single day, accumulating interactions with nearly 10,000 citizens.
As service hours extended, on-site work also increased. In addition to the robot completing rallies, the operators had to organize participants, arrange the order of experiences, and handle equipment operation and maintenance issues.
The significance of pop-up operations lies in testing sports algorithms and on-site service processes within the same environment.
By incorporating personnel organization, equipment debugging, and maintenance into the deployment plan, the next scenario application will have the conditions to reuse existing experience, reducing the investment in re-exploring operational processes.
The 140th Canton Fair further expanded the scenario scope of this practice: the guided tour services and sports interaction plans cover all three phases, and the first phase booth simultaneously showcases industrial transportation, special inspection, and dexterous manipulation solutions.
In public areas, PHYBOT C serves as a "worker" on the Zhujiang River promenade, taking on the roles of human-robot badminton and table tennis rallies and boxing demonstrations. At the same time, it transforms into a cyber guide, providing intelligent guided tour services at the Canton Fair History Museum through autonomous navigation, obstacle avoidance, and bilingual Chinese-English explanations. The above public services and interaction plans will continue until November 4.
The same series of robots undertaking sports interaction and guided tours requires that task models and venue configurations can be switched. When applications expand to industrial and special operations, different embodiments must also be selected based on payload, terrain, and operational requirements.
Focusing on heavy-duty transportation, PHYBOT Technology had the four-legged robot dog PHYBOT D walk on-site carrying large cargo boxes. This robot dog achieves a continuous walking payload of 210kg, with a maximum payload reaching up to 500kg.
The same embodiment can also swap mounted equipment through the development module on its back to adapt to other tasks. In special inspection scenarios, PHYBOT D can be equipped with fire extinguishers, smoke detectors, and LiDAR, providing a reference for application solutions targeting inspection and emergency tasks.
For dexterous manipulation tasks, mobility and load-bearing capabilities alone are not enough; robots need corresponding end-effectors.
PHYBOT Technology's first public release of the fully direct-drive anthropomorphic dexterous hand PhyHand features 23 active degrees of freedom. Aimed at applications such as robotic arm collaboration, precision assembly, and teleoperation, it supplements the fine operation capabilities of the manipulation terminal.
These types of products collectively provide a deployment approach: designing and selecting embodiments based on the requirements of intelligent tasks, and completing adaptation on existing hardware and control capabilities.
When replicating to the next scenario, the team has the conditions to reuse existing embodiment and equipment configurations, focusing development efforts on model adaptation and on-site debugging for new tasks.
The more parts that can be reused, the fewer links need to be redeveloped, giving multi-scenario applications the opportunity to reduce the implementation cost of individual projects.
However, whether technical solutions can run continuously at new venues also depends on whether equipment deployment and operational processes can be seamlessly connected.
In the sports business, PHYBOT Technology chooses to cooperate with sports brands to connect technology development with venue support.
Recently, PHYBOT Technology and Double Fish Sports completed the signing of a strategic cooperation agreement. The two parties will focus on cooperation in applications such as co-building a robot sports interaction laboratory, event interaction, and sports education.
But to replicate a single cooperation to more venues, deployment and operational experience need to form more standardized and mature solutions.
PHYBOT Technology's launched PhyAgents Hub is precisely the solution proposed to address this issue. Taking sports entertainment and commercial services as entry points, it integrates venue conditions, hardware, models, and subsequent operational maintenance into a single scenario solution.
The goal of this solution is not to have partners clarify supporting conditions item by item when determining robot configurations, but rather to resolve on-site deployment and subsequent operations in advance through standardized solutions, enabling customers to operate smoothly immediately after purchase without being distracted by deployment and O&M time consumption.
The changes in PHYBOT Technology's exhibition participation across the two Canton Fairs also reflect this approach of balancing technology and operations.
In April this year, at the 139th Canton Fair, PHYBOT Technology's robots still focused on showcasing badminton intelligent decision-making and interactive experiences.
By the 140th Canton Fair, its robot application scope has expanded to public services, bringing continuous operation demonstrations with different embodiments and multiple scenarios.
Supporting these is PHYBOT Technology's overall coordination of environment adaptation, hardware configuration, model deployment, and operational maintenance. Its full-chain standardized deployment capability has taken initial shape and continues to mature.
04.Conclusion: With One More Scenario, Can There Be Less Repetitive Investment?
Multi-scenario, normalized on-duty operations are a necessary path to commercialization. However, what truly affects the commercialization process is how much new training, debugging, and maintenance work a company must invest in for each additional customer.
Therefore, for companies, the more existing experience can be put to use, the more opportunities there are to amortize R&D investments across more projects.
From this perspective, the scenario demonstrations at the Canton Fair are merely the starting point of the business. How much existing experience a company can apply to the next scenario, allowing customers to spend less time debugging and get into use faster, will determine whether its robots can enter more venues and stay for long-term work.
What is needed behind this is precisely a standardized solution like PhyAgents Hub that unifies venues, hardware, models, and operational maintenance. PHYBOT Technology's existing attempts show that someone in the industry has already taken the first step.
References:
[1] Open X-Embodiment: https://deepmind.google/blog/scaling-up-learning-across-many-different-robot-types/
[2] Gemini Robotics-ER: https://deepmind.google/blog/gemini-robotics-brings-ai-into-the-physical-world/
[3] GPT-6 Astra: https://arxiv.org/html/2609.24170v1