JD has also joined the "poker table" of robots.

At a high-level meeting of JD.com, He Xiaodong, the senior vice president of the group and the deputy dean of the Exploration Research Institute, once walked into the meeting room with a robotic dog. This is not an ordinary robot. As it followed He Xiaodong around, it interacted and chatted with the people present in natural language.

After the meeting, many senior executives gathered around to have a look at this “talking and moving” robotic pet. Some even said on the spot, “I also want to buy one.”

Behind this seemingly relaxed “dog-walking” scene is a signal that JD.com wants to make its mark in the AI and robotics field. At the 2025 World Artificial Intelligence Conference (WAIC), JD.com officially launched its embodied intelligence platform brand “JoyInside”, reached cooperation with many robotics companies, and not long ago, announced investments in three robotics companies in one go – Qianxun Intelligence, Zhujidongli, and Zhongqing Robotics.

In 2025, embodied intelligence became the hottest new battlefield in the field of artificial intelligence. At the just-concluded WAIC, the combination of large models and robots became the main theme, and the exhibition site almost turned into a robotics skills show. Major technology giants showcased humanoid robots, quadruped robots, AI toys, and intelligent agent systems, trying to seize the next wave of AI implementation.

However, in He Xiaodong's view, embodied intelligence is not as simple as “adding a large model to a robot”. It is a complex systematic project involving the “brain (intelligence) + cerebellum (movement) + input-output interaction”. It is much more difficult than expected for it to truly move from the laboratory to family and daily life scenarios.

“The public has high expectations for AI, so it can easily explode at one point. But for it to truly be implemented, such as entering families, there are still many problems to be solved. Large models themselves have not been fully integrated with robots,” He Xiaodong told Jiemian News and other media.

But he is personally optimistic about the industry. “Once any industry starts to reach users, especially C-end users, the iteration speed will be very fast. Just like autonomous driving, robots may follow a similar path.”

The ‘Trilogy’ of Embodied Intelligence

If we were to summarize embodied intelligence in one sentence, it could be “an intelligent physical machine”. It could be a bipedal humanoid, a quadruped dog, or take other forms.

He Xiaodong has a broad understanding of embodied intelligence, not limited to humanoid robots. Any intelligent entity that can interact with users in the physical world and has a certain level of emotional and intellectual quotient belongs to embodied intelligence. However, it needs to have the following three core modules:

First, the “cerebellum” that controls movement, ensuring that the machine can move in the physical world. “Without any movement ability, the experience will be significantly lacking.”

Second, the intelligent central control as the “brain”. Movement ability alone would be monotonous, and the “brain” can enhance the experience, interactivity, and perception.

Third, IO input-output interaction, which is often overlooked but is a crucial part. Embodied intelligent robots involve interactions among the robot, the world, and the user. The user's commands and needs need to be perceived, and the robot also needs to be able to give feedback.

“Many robotics companies in the market are good at control and movement technologies but lack capabilities such as large models, voice interaction, and conversational intelligent agents.” Based on this understanding, JD.com positions JoyInside as an “embodied intelligence” platform – “embodied” on top of these companies to make up for their shortcomings.

Currently, JoyInside has reached cooperation with many companies, including Fuzozo, Huohuotu, Yuanluobo AI, Xin Dalu, Folotoy, Dr. Lukaka, Zhongqing Robotics, Deep Robotics, Magic Atom, etc., and has opened free access to large model technology for a limited time.

Different from most companies that start with industrial robots, JD.com's strategic focus on embodied intelligence is beginning to shift towards daily life scenarios such as family and companionship. The reason is that these scenarios not only have a large scale and high frequency but are also the directions with the most user stickiness and consumption potential. They can better verify the emotional connection ability between intelligent entities and humans.

Take the robotic dog jointly developed by JoyInside and Deep Robotics as an example. This quadruped robot, originally used for pipeline inspection, is now being transformed into an interactive and companionable family pet. It can walk autonomously, chat and interact, and also has a certain ability to perform tasks. He Xiaodong calls it a “walking search engine” or “the prototype of a family assistant”.

“The space in this area will be one to two orders of magnitude larger than before, and the experience for users will be more direct,” He Xiaodong said.

He even compares the popularization of embodied intelligence to the evolution of cars: “Just as every family has a car now, in the future, every family may have a humanoid robot or a robot in other forms.”

In addition to technological platform and product cooperation, JD.com is also accelerating the construction of the embodied intelligence ecosystem through investments. Just one week before the WAIC, JD.com successively announced investments in three robotics companies, Qianxun Intelligence, LimX Dynamics, and Zhongqing Robotics, continuously increasing its investment in the industrial side.

Regarding the investment strategy, He Xiaodong said at a media communication meeting that JD.com values three criteria: First, embodied intelligence is a strategic track worthy of long-term investment; second, it can be integrated into JD.com's service ecosystem and bring a leap in user experience; third, as long as it conforms to the system logic of “brain + cerebellum + IO”, it can be regarded by JD.com as a cooperation and investment target. He also revealed that the next investment plan is still in progress.

Problems in Technology Implementation Need to Be Solved

Embodied intelligence is undoubtedly one of the hottest tracks this year. According to IT Juzi data, there have been more than 110 investment events in the domestic embodied intelligence industry this year, with a total investment amount exceeding 12 billion yuan.

At the just-concluded 2025 WAIC, robots were undoubtedly the focus. Various robots staged a “skills competition”, from rolling and jumping to precise operations. The technology demonstrations attracted a lot of attention. Behind this popularity is the industry's high expectations for embodied intelligence, but it has also sparked discussions about “whether the concept is overheated”.

He Xiaodong believes that the public's high expectations for AI technology can easily make a certain field “explode” quickly. But for embodied intelligence to be truly implemented – such as entering families – there are still many problems to be solved.

“The most typical example is hand operation,” he said. “Although the hand is small, it has many joints and is much more flexible than the foot. Many robots today can do forward and backward rolls, but in terms of the difficulty of degrees of freedom, it may not be as complex as the operation of a hand. For robots to provide services, precise hand control is an unavoidable hurdle.”

Another major challenge is the in-depth integration of large models and robots. For example, spatial intelligence, enabling the robot to quickly identify who is in what position after entering a room; and semantic understanding, enabling it to accurately locate which water cup to pick up when asked to bring the water cup on the table – these seemingly simple scenarios involve multiple technologies such as visual recognition, semantic reasoning, and spatial modeling, which have not been fully solved yet.

Despite the challenges, He Xiaodong remains optimistic about the industry's prospects. He takes autonomous driving as an example. In the early days, there were debates about whether to start with L2 or L3 and whether to use simulated data or real data. But Tesla first launched L2 and collected a large amount of real data through cameras, and there has been obvious technological progress in the past two years. Robots may follow a similar path – first be implemented in specific scenarios, accumulate user data, and then iterate quickly.

In his view, robotic pets targeting the C-end market may be the first track to explode. “Keeping a real dog is costly and troublesome. You have to feed it, take it for walks, and deal with its hygiene. A high-emotional-intelligence robotic dog can provide a similar sense of companionship without the hassle, so the market potential is huge.”

In the process of embodied intelligence entering the C-end market, how to control the token consumption cost and achieve commercialization has also become a common problem for enterprises.

Sun Zhaozhi, the founder of Luobo Intelligence, the development enterprise of Fuzozo, said in an interview with Jiemian News and other media that since the product was launched, users have an average daily usage time of 1 - 2 hours, and the token (the unit of large model call cost) consumption is extremely high. “This type of product is not a simple conversation machine. It has a complex agent mechanism and the ability to manage emotions and memories. Each interaction involves calling multiple models.”

“Currently, no company is making money by selling tokens,” He Xiaodong said. “If it's just a competition based on the token unit price, it will only lead to lower prices and worse experiences, creating an unhealthy ecosystem, just like the price war of early smart speakers, where everyone finally put the experience aside.”

In terms of the pricing mechanism, Fuzozo will adopt a “monthly paid membership system” instead of a one-time purchase. “We hope to create a model similar to that of American SaaS companies.”

Currently, although the trend of embodied intelligence has arrived, there is still a long way to go from a popular product to a truly usable system.

This article is from “Jiemian News”. Author: Zha Qinjun, Editor: Wen Shuqi. Republished by 36Kr with permission.