After looking at Tianmei's layout of AI, I feel that it has really clear ideas.

The power of combining games and AI may be far beyond your imagination.

文/ISAAC

Recently, the World Artificial Intelligence Conference (WAIC 2023) was held in Shanghai. Needless to say, the gold content of this conference is a large number of Turing Award and Nobel Prize winners, domestic and foreign scientists, and many leading AI companies participating in the conference. After seeing some of the developments at the conference, Mr. Putao was really moved.

Because the last time I participated in WAIC was in 2021, the momentum of AI was not as prosperous as it is now, and my understanding of AI was not very good. But this year - the scale of my country's AI core industry has reached 500 billion yuan, the scale of computing power ranks second in the world, and the number of companies exceeds 4,300... There were more than 30 large models displayed at the conference alone. The AI ​​era is approaching unstoppably.

At the conference, the game-related venues were also quite lively. This is not surprising, after all, the gaming industry is one of the industries most closely related to AI. But judging from my experience in this aspect during the period when AI became popular, it seems that solutions that can bring great disruption to gameplay have not yet been implemented on a large scale in the industry. Regarding the future direction of games × AI, there are also different opinions in the industry, and it is difficult to reach a conclusion. Therefore, most people seem to have entered a period of AI confusion—they understand that this path is the future, but they don’t know exactly how to go about it.

Of course, at this time, game companies are indispensable to speak out. This year’s conference opened the “Game AI Application and Game Technology” forum for the first time. Guests at the conference shared their thoughts and practices on games × AI. Among them, Tianmei’s sharing makes me find it very interesting, because in the context of a somewhat confused environment, they seem to show a relatively rare systematic thinking about games × AI by Tencent as a major manufacturer.

If you briefly abstract what Tianmei shared, you can roughly understand it this way: their game × AI planning is divided into two parts: internal and external - "internal" refers to using AI-related technologies to improve the in-game experience, and "external" It is to use games to feed back the development of AI and other technologies. If we can cultivate both internal and external aspects, the two will complement each other and form a good closed loop.

To put this matter into a more specific perspective, you will find that in terms of understanding and application of AI, Tianmei has combined AI to bring about qualitative changes in many game contents, and has also developed many practical references. technical direction.

01

To what extent can AI improve the gaming experience?

First of all, in terms of using AI to improve the game experience, what we are currently discussing most frequently is AI NPC. But in fact, achieving a more complete experience in this regard requires very long-term iterations. In contrast, if you combine the needs of projects and players and focus on some lower-level scenarios, you will find that AI applications have great potential.

A typical case is the application of Tianmei J3. They focus on FPS and have developed games such as "Reverse War", "Cross Fire: King of Guns" (CFM), and "Call of Duty Mobile" (CoDM). Counting the "Reverse War Mobile Game" under development, they have already used AI technology in many projects, and can be regarded as one of the first teams in the industry to explore the development and application of FPS AI technology.

Of course, they will probably be a little confused at first, because FPS AI is a "big pit", and its development and application are very difficult - the environmental factors, status analysis, and situation judgment in FPS... although complex, are very difficult for people. But for AI, the complexity brought by these factors is several orders of magnitude higher than that of 2D/2.5D games. Not to mention achieving a higher degree of completion under various advanced rules and gameplay.

Under this premise, it is very important to find the right research direction for FPS AI. Therefore, after considering the gaming demands of different types of products, J3 mainly sorted out two general directions:

The first direction is the "anthropomorphic AI" of FPS - in layman's terms, it is "human-machine battle".

But unlike the human-machine model we have understood in the past, this kind of anthropomorphic AI is to improve the competitive experience. In hard-core competitive games like FPS, what players need most is to constantly challenge and break through their abilities in a balanced dynamic game. Starting from 2018, J3 began to cooperate with the company’s artificial intelligence laboratory, Tencent AI Lab, and plans to Taking "anthropomorphic AI" as the direction, create some extensions of gameplay and experience.

Technically, the primary difficulty is how to overcome the complexity brought by the 3D environment.

In addition, in order to improve research and development efficiency and enable AI to adapt to multiple modes, in addition to optimizing performance, they also designed a joint training framework for OneModel. Through methods such as generalized shared modeling of the map environment and joint training of multiple maps, AI has the ability to quickly cover new maps.

After solving this problem, we have to face the difficulty of AI training. Their approach is to start researching and training the model from the simplest 1V1 mode, and then slowly conquer the complex mode of 5V5. With the cooperation of the joint team, Tencent AI Lab and J3 jointly implemented the industry's leading FPS full-image anthropomorphic AI solution.

This technology has even become a breakthrough point for products to break through.

In 2020, CFM plans to develop an interactive story mode similar to Gal Game to provide some companion experiences in addition to PVP, while also taking into account competitive content. How to do this kind of "need and want" sound? They have launched a new version of "E-Sports Legend" - based on real-life professional players, using anthropomorphic AI to learn their game data and operations, allowing players to have a feeling of interacting with different characters and challenging players.

Once this mode was launched, it became the mode with the highest retention since the BR mode update of "Desert Island Training" in 2017. Since then, CoDM has also launched the "E-Sports Challenge" mode in 2022, incorporating the main players of the Q9 team that won the Professional Masters Championship that year into the game. By collecting and analyzing a large number of players' real game data, the team customized "AI clones" of professional teams so that players can play against them in 5v5. After the gameplay was launched, it was unsurprisingly popular - the "E-Sports Challenge" record ranked in the top 8 among the game's more than 60 historical modes.

The second direction is "AI action generation" for PVE - the reason why we aim at this direction is because "Reverse War Mobile Game" is an FPS PVE game, which contains a large number of Boss performances, battles, and mobs. Scenes of same-screen battles and NPC character actions are all a very important part of the game's immersive content and combat experience. Nowadays, in order to achieve better results, it is obviously necessary to apply some more cutting-edge technologies.

For example, stacking huge behavior trees and animation resources is not impossible to produce realistic content, but it will consume a lot of energy, and also has the disadvantages of taking up a lot of memory and high maintenance costs. Of course, in recent years, overseas 3A manufacturers have also developed more advanced animation technologies such as MotionMatching and HyperMotion, which have solved some of the pain points of traditional solutions.

And Tianmei’s direction is another way: at the end of 2019, “Reverse War Mobile Game” cooperated with Tencent Robotics X Laboratory and set its sights on the autoregressive neural network algorithm model (ARNN model).

This is a data analysis algorithm widely used in time series, such as weather, traffic prediction, etc. Character animation is essentially a kind of time series data. If this algorithm can be used reasonably, it may be possible to train the AI ​​model through a large amount of motion capture data to learn to predict the next frame of the action sequence frame and form a natural, Full character action.

This technology has been explored by some 3A manufacturers, but it has not been implemented on a large scale in the industry, and few people have set foot in the field of online games/mobile game products. Therefore, the application of "Reverse War Mobile Game" should be regarded as the first attempt of the ARNN model in an online game.

The action reactions of NPCs using this technology in the game will be generated in real time based on the player's actions, rather than animations written in advance; and they can independently generate and adjust realistic action forms based on kinematic principles. . It may not be intuitive to say this. The J3 team demonstrated to us a technical effect - you can see that in the multi-mobs chasing effect animation using this technology, the AI-generated mob character movements are performed in real time. Compared with traditional effects, it is more vivid and diverse, and can also make various changes based on the environment and terrain.

Same screen under intelligent agent action generation technology

Multiple monster chasing effect indication

What does this effect mean? If it can be applied on a large scale, players will be able to see more natural and immersive realistic action performance and combat effects in the game, and this technology will continue to be used on some NPCs and monsters. In addition, the production efficiency will also be improved in this way - for example, content that takes at least two months to complete in the traditional art workflow can be completed in two weeks through AI action generation. On the other hand, this also improves the playability of the game. After all, the team can devote more energy to the production of richer content.

02

Game × The power of AI, may befar beyond your imagination

Of course, the chemical reaction between games × AI goes beyond that – after improving the in-game experience, these AI technologies can also be extended to all aspects of reality through games, achieving feedback effects. At least this is true from these three angles:

First, for more vertical and technology-oriented scenarios, games can become an excellent testing ground for new technologies. We mentioned above that the research and development of FPS AI is very complex, but it is precisely because of its complexity that its value is great - the 3D environment in FPS is more like the real world, and compared to 2.5D/2D AI training, it is more It can be close to human performance in real environments, so there are more opportunities to provide cross-field research value.

For example, the AI ​​action generation technology used in "Counter War Mobile Game" can complement each other with the robot research in the Robotics It behaves like a real creature; on the other hand, it can also use the virtual environment of the engine to accelerate the "leveling up" of the robot in the game. In fact, the cooperation between "Counter War Mobile Game" and Robotics

Tencent independently develops the quadruped robot Max

Second, for a wider range of life-oriented scenarios, games × AI can make our lives better. For example: Tianmei Health, a sub-brand of Tianmei, developed a "Tianmei Health Intelligent Fitness Magic Mirror" based on AI technology. This magic mirror was also displayed at the conference that day. At the technical level, it uses AI motion capture and other technologies to provide sports data analysis and result feedback, and can help users formulate scientific training plans and evaluations. Coupled with the combination with game IP content, it can make the originally boring sports more intelligent and interesting.

Third, gaming × AI can even boost the development of the entire industry in the future. We have just talked about that the AI ​​industry and education actually have a big pain point right now - talent anxiety. This emerging field has only been developing for a few years, and a particularly complete and established teaching system has not yet been formed in China. At the same time, more people need to have a real interest in it instead of studying it for temporary benefits. These problems may not be solved naturally in a short time.

When games are combined with AI technology, it can be of considerable help. For example, the open AI research platform "Tencent Enlightenment Platform" jointly created by "Honor of Kings" and Tencent AI Lab has cooperated with 19 top universities such as Peking University, Tsinghua University, and the University of Science and Technology of China to develop courses to allow students to understand in a more interesting way Machine learning, reinforcement learning, multi-agent decision-making and other knowledge points. At present, enlightenment-related courses have been offered to a total of 630 students. Not long ago, they also announced that the AI ​​competition held on the platform will soon be held open nationwide.

If these courses, materials, and competitions can be promoted to AI majors in colleges and universities across the country, they will undoubtedly bring great benefits to our AI education and even change the lives of many students. This is why I say that the power of games × AI is far beyond your imagination.

03

What is Tianmei’s understanding of games × AI?

With this idea of ​​combining games and AI both internally and externally, Tianmei's overall application of AI has reached a very efficient level.

For example, in the sharing by Lin Zhichao, the program leader of Tianmei T1 Studio, he said that there is an open world action game under development, and almost all aspects of development are combined with AI technology to improve the development process. This is a natural thing, because the open world inherently requires rich action and narrative resources to support its free and open characteristics. AI can often solve many development pain points.

For example, in terms of lip-sync animation generation, they applied a generation scheme based on rules and deep learning to extract phoneme sequences from the audio, convert them into visual element sequences, then generate corresponding visual element curves and convert them into action unit curves. Drive animation. This approach can better adapt to different languages ​​than traditional solutions, and can also achieve real-time interaction and customized voice.

For example, in the illustration generation pipeline, they also combine AI painting to create narrative plots and improve efficiency. Usually the team will first draw the manuscript, use AI to draw the characters, and then perform style transfer and manual refinement. Finally, a high-quality original painting that can be put into use can be completed in less than 2 days. In addition, they will also train art resources through models and control details through ControlNet to facilitate subsequent mass production.

With this kind of thinking and application, it can be seen that AI is not so difficult to integrate with game development, but has truly changed the game development model and penetrated into various pipelines. This is why I said that Tianmei has done many big things without knowing it:

First of all, they really "dare to use". When many people are still confused about AI, applying it on a large scale in projects is a very advanced thing. This requires not only excellent technology, but also corresponding cognition. For example, technologies such as AI automatically generating actions and animations require breaking the previous bulk thinking.

Secondly, they have enough patience. Many technologies take at least three to five years to try out after deciding to develop them. During this time, the apparent benefits are likely not to be high. This shows that investment in AI is often a long-term decision, rather than a decision made for short-term dividends.

Finally, their understanding is deep enough. At present, most of the industry's imaginations about games × AI are dotted or linear. For example, when we see the rise of AI painting, we just talk about the iteration of the art pipeline, but there are not many more in-depth and three-dimensional thinking like Tianmei. With this kind of thinking, I believe that more companies in the game industry will make good use of AI. Our industry’s understanding and utilization of AI will certainly not lag behind others in the future.

Of course, in the future, Game × AI will definitely explode with greater energy. At this forum, Yu Dong, the head of CoDM of Tianmei J3 Studio and the development director of "Reverse War Mobile Game", mentioned two ideas:

The first idea is AI automated testing in the research and development process. The degree of test automation is a very important indicator in the DevOps field, but it is still difficult to implement industrial software-level automated testing for PvE games with complex logic and rich content.

But in the future, when AI gradually has the ability to perceive the environment and atmosphere, real-time analysis capabilities, and the ability to make independent decisions about test boundaries and conditions, perhaps this scenario will come true: input a task, and the AI ​​can play the game like a real person Execute instructions, explore experiences, and generate detailed test and feedback reports. This will greatly improve the quality and efficiency of game testing, ensuring and improving game quality.

The second fantasy is to use AI technology to build a more vivid, autonomous, and real virtual world where every NPC seems to be alive. In this regard, there are already teams that have actually invested in research and developed the prototype of an AI society. For example, this year, researchers from Stanford University and Google created a "virtual town" in which 25 AI agents can "survive autonomously", interact and socialize with other AIs. After the paper on this experiment was released, it aroused great attention and discussion in the industry.

If we want to make this kind of society a more perfect "world", we still need to overcome some problems, such as at least giving AI the ability to have natural language, behavior and decision-making, and social interaction.

Prior to this, our exploration of AI still needs to be down-to-earth and step by step. An essential step in the process is for more teams like Tianmei to devote long-term energy to research, improve the game experience internally, and output breakthrough technologies externally. If more and more teams do this, I believe that the reality of "Westworld" is not far away from us.

Game Grape Recruitment Content Editor,

Recommended reading

Lilith Ten Years | How to evaluate Tencent | MiHoYo Twelve Years

Dragon Fist Storm Defeat | Back to the Future: 1999 | New Products in the Next Three Years

Dialogue with Luo Zixiong | Dialogue with Ma Xiaoyi | Dialogue with Cheng Liangqi

Set as star (The process is as shown above)

You can receive push notifications and complete covers as soon as possible!

The above is the detailed content of After looking at Tianmei's layout of AI, I feel that it has really clear ideas.. For more information, please follow other related articles on the PHP Chinese website!