Desktop robots are becoming a new track for embodied intelligence—will the endgame be physical Agents?

In the past two years, on many occasions, especially at various tech exhibitions, we have often seen all kinds of humanoid robots. They are often the focus of these exhibitions, attracting countless visitors to gather around and take photos.

However, although we have always believed that humanoid robots entering the home is an inevitable trend, at present their retail prices are still on the high side, and there is still a fairly long road ahead before they truly enter ordinary people's homes. It is precisely against this backdrop that a type of small desktop robot, designed for office desks or study desks, has gradually come into our view and attracted the attention of some users.

Of course, views on desktop robots are also fairly polarized. Those who like them believe they are the entry-level carrier for large models to land in the physical world; those who find them gimmicky believe they are purely expensive electronic figurines wrapped in the cloak of AI, and once the novelty wears off, they will inevitably end up gathering dust.

So, will desktop robots become a hot track in the field of embodied intelligence?

Desktop scenarios have become a new branch for embodied intelligence manufacturers

Before formally opening this topic, it is still necessary to briefly discuss the definition of desktop robots. At the most basic level, we believe a desktop robot must at least meet the conditions of embodied intelligence + desktop scenario. Then, a desktop robot needs to have AI proactive perception capabilities and must possess a brain, which can come from the computing power provided by a local chip or from the cloud; in addition, since it is a robot, it must also have a mechanical structure device. Normally, we would not classify AI speakers or AI cameras as desktop robots.

Compared with full-size humanoid robots, desktop robots have lower manufacturing costs and a narrower range of capabilities, which is obviously an advantage for startups.

As mentioned at the beginning, one of the major pain points of humanoid robots at present is price. The high prices of full-size robots today ultimately come down to the fact that costs have not yet been thoroughly brought down. Normally, a humanoid robot has a built-in high-performance chip to support local computing power; at the same time, it also needs a brain composed of a large model + agent, which tests algorithms and training data behind the scenes; in addition, the robot's dexterous hands, mechanical structure, and so on are also unavoidable costs.

By contrast, commercialization in desktop scenarios is much easier for robots. The physical boundaries of a desktop are very clear, and the environment is relatively simple, since the range has already been defined: a user's desktop with limited area and space. As a result, it does not need to solve complex bipedal balance problems, nor does it need global navigation and obstacle avoidance; basically, it only needs to do a good job of preventing falls at the edge of the desktop. This means the R&D team can focus limited computing power, battery, and other hardware resources on intelligence and interaction.

Moreover, from the perspective of the supply chain, desktop robots also have cost advantages. Their main control SoC, display screen, microphone, camera module, battery, and so on can almost all directly use the mature solutions accumulated over more than a decade in the smartphone and IoT industries. As a result, component procurement costs are much lower than those of full-size robots.

Not long ago, we introduced in an article that UCLA launched a low-cost bipedal robot solution. Specifically, it uses second-hand phones to drive small robots, keeping the body cost within 3,000 yuan. The core reason its cost can be pressed so low is that it makes full use of various components on the phone: the camera for visual recognition, the SoC for inference computing, WiFi for networking, and the battery for extending endurance. Robots of this size are very suitable for desktop scenarios. Conventional desktop robots will certainly not directly use phones for driving, but many components can indeed reuse the smartphone supply chain.

(Image source: PhoneBot)

There is also a very practical reason many startups enter the desktop robot field: full-size robots are the technical high ground in the industry, and a large number of heavyweight players have already rushed in, making competition fierce, with most still in the stage of continuously burning money. By contrast, the technical threshold in the desktop robot field is relatively lower, the cost of trial and error is not as high, and the demand for capital is not as great. More importantly, it allows manufacturers to quickly launch a product that can be sold directly to C-end consumers and continuously generate cash flow.

Desktop robots focus on emotional companionship and efficiency tools

When we turn our attention back to the products themselves, we find that the main functions of desktop robots currently on the market fall mainly into two categories, namely emotional companionship and efficiency improvement.

First is emotional value and companionship, an attribute many robots now emphasize. In our view, the emotional companionship of desktop robots is mainly presented in the form of a desktop pet that can move, speak, and interact.

For example, the currently popular Letianpai desktop robot is a small bipedal robot whose head is mainly a circular screen, usually displaying two bouncing eye patterns, positioning itself as a cute pet.

The Letianpai desktop robot has a rich variety of interaction methods. In addition to touch and voice control, it can also recognize users' gesture commands through the camera. In terms of movement, it can walk forward and backward and also move left and right. Frankly speaking, it has no particularly explosive functions; its core selling point is emotional companionship, as it can chat and provide feedback through actions. It runs on the Android system, supports flashing Linux, and allows custom large models, offering fairly high playability for advanced users.

(Image source: Zhihu @Gao Baifei)

What makes emotional companionship even more thorough is the Lingtong Nian NIA-F01 humanoid robot. It is essentially a figurine robot, and its appearance is entirely that of a figurine type. The only difference is that it adds a camera, a micro-skeleton, and other sensors. Unlike traditional figurines, it can perform different actions according to commands and can also interact with users via voice and other forms. Users can also create and edit its persona, voice, and so on themselves. Put plainly, it is a highly customizable figurine, and it can be said to have truly mastered emotional companionship.

(Image source: JD)

Next is efficiency tools. Among them, the myCobot series launched by Elephant Robotics is relatively representative, and its form is closer to a small robotic arm. myCobot is not a consumer-grade product that works right out of the box; it requires secondary development to be used. For example, it can be equipped with grippers and suction pumps to become a tidying robot capable of cleaning up a desktop; it can also be connected to a mouse and keyboard to simulate a person operating a computer and help users complete a large amount of repetitive work.

(Image source: Elephant Robotics)

There are more and more desktop robots, mainly because the underlying technologies supporting the corresponding scenarios have already become relatively mature. Now, multimodal large models allow them to understand complex contexts and human natural language. In addition, with the spread of machine vision technology, robots can use front-facing cameras to capture users' faces and gestures and understand them; in terms of movement, desktop robots from major manufacturers basically all use micro servo motor solutions, allowing the desktop robot body to make anthropomorphic movements such as shrugging, shaking its head, and stomping its feet.

Looking at markets at home and abroad, many brands have entered this field. Domestically, Letianpai is a brand that has moved relatively quickly and has done quite clear product definition. They deeply integrated a large model into a compact body, enabling it to chat and make expressions, and entered the market directly at a price of 1,799 yuan. In addition, brands such as Lingqi Technology and Kechu Weilai are also making efforts in different forms, and some major manufacturers that make humanoid robots, such as Unitree, Lingtong, and Xuanji, have also come to make this type of product, such as the figurine desktop companion robot NIA-F01 mentioned earlier.

In overseas markets, crowdfunding platforms are the main battleground for this type of product. For example, the previously popular Looi Robot had a product manager with a bold idea who directly used the user's smartphone as the robot's brain and face, with the base only responsible for power and support, pushing the price down to $129 and raising a very high crowdfunding amount. In addition, the AI open-source community Hugging Face also launched the desktop robot Reachy Mini, mainly aimed at developers.

Overall, there are currently more domestic brands entering the desktop robot field, while overseas brands are still at a small-scale stage.

Is the endgame of desktop robots a physical Agent?

To be honest, the main selling points of desktop robots currently on the market for ordinary consumers are still concentrated on emotional companionship. There are, of course, many desktop robot solutions focused on efficiency tools and practical functions, but many are aimed at commercial use and often only provide a solution such as a robotic arm, requiring secondary development. Although customizability is strong, apart from industry users, basically only geek players can handle them.

Emotional companionship is certainly an important track in the robotics industry, but the desktop scenario itself has certain limitations compared with all scenarios. If emotional companionship is the only main selling point, then desktop robots can only become cyber pets under the banner of embodied intelligence. After all, for ordinary users, it is just a smart speaker with a screen that can walk a couple of steps. Moreover, even entry-level desktop robots often cost more than 1,000 yuan, more expensive than smart speakers.

In our view, if desktop robots are to truly become a track with huge incremental growth, their product form and core functions must undergo a key upgrade.

In terms of product form, desktop robots should move toward modularization. Current products still look too much like abrupt electronic devices. Barring surprises, future desktop robots will deeply integrate with users' monitor stands, desk lamps, and even wireless charging docks. When users need it, it can appear in time and become an assistant; when not needed, it can quietly hide away.

Moreover, modular design will give it greater plasticity. Users may be able to assemble it like a computer, installing a robotic arm module to organize and store small items, or swapping in a tracked chassis or bipedal legs to let it move on the desk.

In terms of functional evolution, desktop robots need a major upgrade in their capabilities in the physical world. Current robots, although they look quite impressive in demonstrations, often run according to preset programs and are essentially script tools. We believe that the true value of desktop robots to users must ultimately come down to doing work. At least for me personally, an ideal desktop robot could operate the computer, organize items, adjust lighting, and so on on my behalf, that is, evolve into a "physical Agent," capable of doing at least more than the AI assistants on phones and computers and AI speakers.

Overall, although current desktop robots are riding the heat of embodied intelligence, they are still in the early stage of industry development. Consumer-grade products are mainly focused on emotional companionship and have not yet truly brought out their ability to improve productivity.

In our view, desktop robots should not be just expensive figurines that can move or smart speakers with eyes. Since they occupy the user's precious desktop space, they should provide efficiency value that matches it.

Barring surprises, with the spread of multimodal capabilities of large models and the continued reuse of components from the mobile phone supply chain, the desktop robot track will see more new products in the coming years, but for them to truly make ordinary users willing to pay, they still have a long way to go.

This article comes from the WeChat public account "Leikeji," author: Leikeji, published by 36Kr with authorization.