Will industrial robots become the first stop for large models? How do large models empower industrial robots?

Humanoid robots have attracted a lot of attention in the industry in 2023, and capital is obviously fond of them. However, since there has not been a real landing, or even a normal operation demonstration, a large number of people have not easily entered the market, but are waiting to see whether the pioneering companies can make landing products, and then follow up. Therefore, rather than setting off a wave of humanoid robots, it is better to say that thunder is thunderous but raindrops are small.

Looking back at the end of the year, what can truly be called the biggest trend in the industry this year is undoubtedly .

Judging from the direction of capital, artificial intelligence can be said to be almost the only bright spot in 2023. Due to the sluggish global economic situation and the influence of geopolitical factors, capital has entered a cold winter period.

According to data from tchBook, an investment and financing analysis agency, the total amount of global venture capital in the first half of 2023 was only US$173.9 billion, a year-on-year decrease of 48%, and the number of financings also decreased by 19%. In the third quarter, the situation did not improve, and the global financing amount decreased by 31% year-on-year.

However, artificial intelligence has risen against the trend. In the third quarter, global financing for artificial intelligence companies increased by 27% year-on-year to US$17.9 billion, even exceeding the total financing of other hard tracks.

It is no wonder that a certain capital industry practitioner said: "This year there are only two technology tracks, and others."

From the industry's perspective, although large AI models represented by Chatgpt are in the limelight, their application scenarios remain a mystery. Since large models currently perform well in the fields of text and image generation, many people believe that this wave of AI is more likely to be implemented first in the service or design industry.

But in fact, Gaogong Mobile Robots found that many mobile robot companies believe that artificial intelligence is a key technology to promote the explosion of mobile robots. They are currently integrating the progress of artificial intelligence into their product systems , and several of these companies emphasized that "this is definitely not a gimmick."

An industry insider said: "You said there are no bright spots in the industry this year, but I disagree. Since the beginning of the year, AI has given us a lot of inspiration. I think this is an opportunity for the industry to evolve and a change that can really lead to the explosion of mobile robots."

Perhaps the first industry to be impacted by this wave of artificial intelligence revolution will be industrial mobile robots.

0****1

How do big models empower industrial robots?

After being shocked by the revolutionary capabilities of big models at the beginning of the year, people in the industry have now returned to rationality. The specific application scenarios of AI big models and how much profit they can bring have become the core of industry concern.

According to the visit of Gaogong Mobile Robot, artificial intelligence is currently mainly used in three areas:

1. Human-machine interaction . The control, scheduling and troubleshooting of robots have always required professionals who are familiar with relevant knowledge, which means high labor costs. Due to the amazing progress of big models in natural language processing and generation, and other fields, with the support of artificial intelligence, on-site staff can directly use natural language to let big models generate codes to control robots. Therefore, the interaction with robots can achieve low-code or even code-free human-machine collaboration.

At the same time, unlike traditional human-computer interaction methods, operators can input multimodal instructions (such as pictures, text and language) and correct the robot's actions through multiple rounds of commands.

In fact, this is also one of the main ways that technology giants try to implement big models. In March, Google launched PaLM-E, which focuses on using natural language to command robots to grab rice cakes; in April, Alibaba revealed that it was experimenting with connecting the Qianwen big model to industrial robots, using natural language to remotely command robots to work, and released actual operations.

**2. Vision. **Industry insiders said that although image recognition is one of the most important application areas of artificial intelligence, it was not until this year that the application of AI+ in the field has made revolutionary progress. According to them, in the past, when applying artificial intelligence, they needed a large number of image annotations and various defect samples, but there would still be new defects that had never been seen before, and the computing power for training, the time for deployment, and the labor cost required were all very high.

Now, with the emergence of large models, they only need to provide good samples to complete the deployment on the production line in a shorter time with minimal computing power. Similarly, compared with the currently commonly used technologies, the mobile robot's ability to identify the surrounding environment and avoid obstacles in complex indoor and outdoor scenes such as strong light has also been improved.

3. Embodiment . With the development of large models, the robot's ability to understand and process multimodal information has been unprecedentedly improved. As a multi-sensing integrated device, the robot can obtain information through vision, touch and other means at the same time. Therefore, after visually identifying a certain raw material, it can adjust the gripping force or processing action according to the material.

It is worth mentioning that if we observe the development of artificial intelligence and the robotics industry calmly, we will find that the integration of industrial robots and artificial intelligence has a long history, and many companies have also explored multiple paths for implementation.

The head of a well-established robotics company told Gaogong Mobile Robot that they began exploring the application of artificial intelligence in mobile robots as early as around 2017. At that time, they were doing relatively simple human-computer interactions based on language recognition, and could use natural language to command a robot to move to a certain location or perform a certain action.

The big model craze led by Chatgpt this year has not so much completely changed the path of industrial application of artificial intelligence as it has allowed companies to explore these paths faster and with greater imagination.

Of course, this does not mean that the big model is a gimmick. It does revolutionize the speed of deployment, iteration, and evolution of artificial intelligence.

The head of a robot company's product line said: "For example, the big model enables robots to have the ability to analyze intent , which is a huge difference compared to before. In the past, you had to tell it very precisely that it should do specific actions ABCD. But now, it can recognize intent. If you tell it AB, it will mainly ask you whether you want to do CD. When it can recognize the operator's intention, it means that we can use fewer keyword inputs and shorter training time. The robot can make associations and analyses by itself, allowing artificial intelligence to handle it autonomously."

The accumulation of R&D and practical experience over the past years, as well as the continuous construction of digital infrastructure to ensure the operation of artificial intelligence (such as edge computing technology, databases, big data platforms, etc.) have enabled the industry to quickly embrace emerging technologies after the rise of big models. It may become the industry that first implements and benefits from big artificial intelligence models.

0****2

Why has the AI+industrial era not arrived yet?

Although the future prospects are bright, the journey of large models to be implemented in the industry is still bumpy.

This is mainly due to three reasons:

1. Currently, the development of large models is still in its early stages, and most companies are unable to explore it . Gaogong Mobile Robots found that many companies believe that large models will definitely have great potential in the industrial field, and have invested manpower and material resources in research, but they all said that the cost is too high, so it can only be a game for giants. With the current strength of Chinese robot companies, it is almost impossible to build a basic model from the bottom up. "Don't we want to do it? No, we are too weak and too small, there is nothing we can do!"

2. Lack of training data. It is well known that large AI models require a large amount of data for training, but in the industrial field, the digital transformation of many companies is still in the very early stages and cannot provide enough data for AI training. Moreover, data in the industrial sector is a very core resource for each manufacturer and cannot be shared, let alone opened for external large model training. Therefore, AI cannot exert its full strength.

On the other hand, the complexity of the subdivision scenarios in the manufacturing industry also makes it difficult for companies to provide sufficient training data. Even in the same subdivision track, there will be very different scenarios in the same production chain, and the data they generate and the demand for artificial intelligence are also different.

A company leader once said: "Manufacturing is a trillion-dollar market in the long run, and a 100 billion-dollar track in the near run." This also means that either different data is collected for different scenarios, or the same general large model is used in different scenarios. However, the former is too costly and may not have enough data; the latter is difficult to adapt to different scenarios that are very different.