When the word "robot" was first coined, it was not originally intended for performances.
In 1920, Czech writer Karel Čapek coined the term in his play R.U.R. (Rossum's Universal Robots), borrowing from the Czech word "robota," which means forced labor or drudgery. Although this may seem somewhat "inhumane" from today's perspective, in the original play, those artificial humans were mass-produced precisely to do the tiring, tedious, and dangerous work for people.
However, over the past few years, the most familiar image of robots to the general public has been them dancing and doing somersaults on stage. The original duty of "working" hidden in the etymology was once overshadowed by lively performances.
Today, the Ministry of Industry and Information Technology (MIIT) and the State-owned Assets Supervision and Administration Commission (SASAC) have explicitly set the goal of promoting the deployment of 10,000 units by the end of 2026, and relevant policies are being rolled out one after another. People are tired of watching performances, and even if the industry has been slow to react, it must now realize: the time has come for embodied AI to directly face the challenges of practical deployment.
However, what exactly makes these challenges so difficult? And how should they be resolved?
There are two major mountains standing in the way of embodied AI: one is called "cognition" and the other is "execution." Fortunately, they are not immovable. On the occasion of the 70th Canton Fair, we saw the answers on-site.
![]()
The "Brain" That Robot Companies Cannot Build, Supplemented by General Large Models
Let's first talk about the cognitive challenge.
Robots can already do quite a lot: they can walk, run, grasp, and lift. So why are they still a distance away from scaled deployment? The problem still lies in the core "brain." Once tasks are extended and require completing multiple things in sequence, the deep thinking needed to figure things out clearly before executing step-by-step is still lacking.
Embodied AI companies have also worked hard on this, but to solve the problem of deep thinking, the cost and cycle of developing a "brain" in-house are incalculable, and after many years, no breakthrough has been seen.
The turning point appeared on another track: The general large models, which sparked heated discussions some time ago, have demonstrated astonishing logical reasoning and long-horizon planning capabilities in the digital world. This has given the industry a new solution—since it is too difficult to build a "brain" from scratch, why not directly "borrow" an existing general large model? With this in mind, let's look at what has already happened in the industry.
In September this year, OpenAI released GPT-6 Astra. There is not a single word related to robots in its official introduction; it focuses on digital world tasks such as operating computers, writing code, and handling professional work. Yet, such a model that never intended to "be a robot" achieved an average success rate of 22.48% in 42 simulation tasks and 2,100 tests in an evaluation published by the robot benchmark RoboDojo, surpassing all 40 publicly available robot policies at that time.
It is worth mentioning that among these 42 simulation tasks in RoboDojo, many require continuous multi-step operations and long-term decision-making, such as tidying up a desk, throwing bottles into a trash can, sorting a pile of objects, and playing tic-tac-toe. From the results, it can be seen that general large models, without special training, can solve long-horizon tasks and multi-task planning to a certain extent, which also means that the capabilities demonstrated by general large models already meet the conditions to enter robot workflows.
But obviously, just one case is not enough to provide sufficient evidence:
At the end of September, PHYBOT Technology announced three sets of experiments to the public, with the final conclusion pointing to: the collaboration between general large models and embodied AI will push embodied AI from single-point capability breakthroughs to a closed-loop of system-level autonomous capabilities. The understanding of human intent, long-horizon tasks and multi-task planning, and the grasp of logical sequences—these problems that were once thought to be solved by embodied AI companies themselves—have begun to be accelerated through division of labor and collaboration. For the industry, this means the deployment pace of embodied AI may be faster than expected.
Subsequently, on October 12, Wang Xingxing, founder of Yushu Technology, also talked about the impact of general large models such as GPT-6 Astra. He believes that the tipping point for embodied AI could arrive in as little as one to two years, and frankly admitted: "The biggest mistake I have made over the past decade or so is not believing in AI enough." The implication is that the speed of progress in general large models has exceeded everyone's expectations, and the combination of embodied AI and general large models is highly likely to accelerate the arrival of this tipping point.
The results of evaluations and experiments, as well as the judgments of industry insiders, all prove that using general large models as the "brain" of robots is theoretically feasible, and there is a possibility that the "cognition" mountain is about to be crossed. So, can the embodied AI industry successfully take over the "execution"? Perhaps we can see more details at the 140th Canton Fair.
Delivering Results On-Site at the Canton Fair, Robots Move from Booths to Workstations
Let's first briefly review some of the changes we have seen at this year's Canton Fair.
CCTV News summarized the changes in the embodied AI industry at this Canton Fair into three layers: products, functions, and clients. In terms of products, more and more robots are moving from mere display to taking up posts, with explanation and delivery, firefighting, sorting, exoskeleton assistance, and exterior wall cleaning corresponding to different positions respectively. In terms of functions, besides industrial types, companion and service types are on the same stage. In terms of clients, the demand has shifted from mere spectating to negotiating with specific processes in mind.
We will just touch upon functions and clients; our focus is on the performance of "robots collectively taking up jobs." For example, during the Canton Fair, AgiBot deployed its wheeled robot Elf G2 to subway stations leading to the exhibition halls, such as Xingang East Station and Pazhou Station, to provide security check reminders, order guidance, and route inquiries. It can communicate with clients in Chinese, English, Cantonese, and some minor languages, proactively prompting when encountering luggage, children, or wheelchairs. On-site staff retain the right to take over, allowing for timely adjustments in case of equipment anomalies.
Zhidong Weilai also put its own robots to work. Its wheeled humanoid robot "Xiao Zhi" served as an explainer in the CF Award exhibition hall. Relying on the 3D LiDAR at its bottom, it scans maps, fixes positions, and avoids obstacles in the crowd, introducing the exhibits one by one to global clients in Chinese and English.
The robots from PHYBOT Technology mentioned above were also busy with the comprehensive services of the 140th Canton Fair. Slightly different from the previous two embodied AI enterprises, PHYBOT Technology focuses on bipedal humanoid robots, which not only provide services but also bring emotional value. In addition to undertaking intelligent explanations in the Canton Fair Development History Exhibition Hall, its bipedal humanoid robot PHYBOT C series also conducted autonomous badminton and table tennis hitting experiences with on-site clients near the densely populated Zhujiang Promenade.
Notably, these interactions required no manual remote control throughout the process, nor were there any preset fixed programs. They demonstrated a high level of intelligence and stable motion control capabilities, winning continuous praise from experiencers and onlookers, who eagerly wanted to step forward and try a few moves. In the boxing demonstration area on the other side, the C series robots from the red and blue camps were also fighting fiercely on the ring, drawing rounds of cheers.
It is understood that before serving this Canton Fair, multiple capabilities of PHYBOT Technology's bipedal humanoid robots had been repeatedly verified.
In April this year, PHYBOT Technology brought the global debut of the "badminton high-dynamic intelligent interaction" scenario at the 139th Canton Fair; in July, a pop-up store centered on the core experience of humanoid robots autonomously playing badminton was deployed for a month in a Guangzhou mall; in August, the table tennis experience migrated from badminton capabilities also began to make frequent appearances. This series of actions made us notice two details: PHYBOT Technology's continuous operation and cross-scenario migration capabilities.
In terms of continuous operation, PHYBOT Technology has its own methodology.
For robots to enter different venues for deployment and operation, spatial conditions, equipment configuration, model adaptation, and subsequent operation and maintenance must be considered. PHYBOT Technology chose to build standardized scenario solutions on its own, launching PhyAgents Hub, which packages the five modules of environment, hardware, model, operation, and maintenance into a reusable modular scenario deployment kit, saving clients from building a complete robot service system from scratch. This is also the key to PHYBOT Technology's ability to deploy badminton interactive scenarios in places with significantly different environments, such as malls and exhibition halls.
The reason why the technology can be reused lies in the core of the physical agent PhyAgents in PHYBOT Technology's Embodied PhyStack technical architecture. It is a decision-making agent trained through reinforcement learning, featuring strong generalization ability, high robustness, small parameters, and low computing power requirements, enabling the rapid deployment of trained capabilities to the PHYBOT entity. Therefore, the real-time decision-making trained in badminton can be migrated to table tennis with slight adjustments, without starting from scratch for every scenario.
From single-point demonstrations to continuous deployment, the above capabilities are being tested on-site at the Canton Fair.
This is not a pop-up event of just a few days—it is understood that this batch of robots from enterprises such as PHYBOT Technology and Zhidong Weilai will continue to serve until the end of this Canton Fair on November 4. Among them, the stable execution of bipedal humanoid robots in multiple scenarios over a long period is a fact that deserves more attention at this Canton Fair.
This is the solution we see embodied AI enterprises adopting to cross the second mountain at the Canton Fair.
Over the Hill, How to Break Through the "Commercialization Test" of Embodied AI?
Over the hill, there is still the last mile to go for the scaled deployment of embodied AI, and the signpost for this last mile reads three big words: commercialization.
For B-end clients, commercialization is extremely realistic: robots cannot just make a stunning appearance on the booth, let alone having to tear down and rebuild the software and hardware every time they enter a new scenario. If customization costs remain high and deployment cycles are long, enterprises simply cannot calculate a clear return on investment, making the deployment of 10,000 units out of the question.
This requires the industry to deliver a commercialization answer that balances "efficient deployment" and "flexible execution." Around how to cross this threshold, major leading manufacturers have taken different paths.
On the one hand, there is the systematic solution represented by PHYBOT Technology.
In the past, many manufacturers easily got stuck on high marginal costs during commercial deployment, either betting everything on cloud-based large models, resulting in slow responses that cannot meet real-time operations, or only focusing on the hardware entity, lacking the support of a smart brain.
As can be seen from the above, PHYBOT Technology's commercialization logic is to tightly bind software and hardware at the underlying level, enabling deep collaboration between the brain and the body. The core purpose of this systematization is to significantly lower the threshold for robots to switch between different work scenarios, allowing the same underlying architecture to be quickly replicated to more tasks; and by providing a one-stop solution covering the entire process from scenario deployment and equipment debugging to daily operation and subsequent maintenance, it truly helps clients reduce comprehensive deployment and O&M costs.
On the other hand, many peer companies choose to secure a position through differentiated scenarios, gaining a firm foothold in segmented tracks.
For example, Yunshenchu, deeply engaged in special emergency and industrial inspection, relies on robot dogs of various forms to precisely meet the actual needs of rescue and security. This exploration avoids the long cycle of all-functional generalization. By persistently working in vertical scenarios, they can also run through their respective closed loops on the eve of commercial deployment.
From the perspective of the overall industry, whether trying to build a systematic base to reduce cross-scenario reuse costs, or achieving the extreme in specific vertical fields to meet the needs of specific buyers, everyone's choices vary.
It is precisely because of these diversified commercial explorations that embodied AI can truly move from merely "showing off muscles" to the pragmatic stage of scaled replication. The final answer to this test is being accelerated in these practices.
Conclusion
Looking back, the two mountains facing embodied AI now have their respective answers. For the mountain of cognition, robot companies are expected to find the key from general large models; for the mountain of execution, the robots actually working on-site at the Canton Fair, especially the bipedal humanoid robots, have already proven it.
Able to think clearly and act effectively, robots have completed the first step from performing to taking up posts.
But after crossing these two mountains, there is still the last mile to go: whether enterprises are willing to pay and whether the solutions can be replicated to the next scenario are the real questions of this test.
It is also for this reason that the industry has developed different solutions such as systematic bases and vertical scenario positioning. The routes may differ, but the directions all point to one destination: enabling robots to truly step into the physical world, take up posts, and work, helping humans solve work tasks that they cannot or are unwilling to complete.
At the 70th Canton Fair, what we see is precisely the moment when this industry is answering the questions seriously for the first time.