SenseTime’s “StarCraft 2” AI Agent Debuts, Showing SenseCore’s Professional-class Decision-making Capability
Technical articles
Released on2022-06-13 17:50:49
Recently, Mr. Zhou Hang, former Chinese Champion of StarCraft 2 and current Head of Technology Research and Development of SenseTime Smart Decision and Game Business Department

Recently, Mr. Zhou Hang, former Chinese Champion of StarCraft 2 and current Head of Technology Research and Development of SenseTime Smart Decision and Game Business Department

Man-Machine Battle Rekindles Suspense.

In the multiple matches demonstrated, DI-star showedthe capability of Grand Master level - the highest level for human race , and performed as stunningly as global top-class professional players.

Mr. Zhou Hang mentioned, “DI-star has reached the level comparable to professional players in only five weeks of training. Such efficient training results are exactly derived from SenseTime’s leading technical strength in the decision-making AI field and the solid computing power support provided by SenseCore.

At present,DI-star and its decision-making intelligence platform, OpenDILab, have been open-source through GitHub, enabling every interested AI researcher or player to dive into them, explore and practice the development and application of decision-making AI technology, and build a decision-making AI innovation ecosystem.

Fearless DI-star Efficiently Wins

The space complexity of “StarCraft 2” is as high as the 1685th power of 10, way above that of Go (the 170th power of 10). Therefore, this game has become a wonderful stage to test AI’s decision-making capability.

During the live broadcast, DI-star showed its flexible and variable decision-making capability in multiple match demonstration clips, and found the optimal strategy again and again in complex matches.

DI-star established early advantages through fine micro operation.

During the course of training, DI-star has integratedsupervised learningandreinforcement learning, which are part of SenseTime’s cutting-edge technologies.

Through supervised learning, DI-star makes AI repeatedly observe high-quality videos of matches between professional players, learn how to play like human players, and quickly master high-level decision-making behaviors.

On the basis of supervised learning, DI-star enables AI to play more than hundreds of millions of matches through self-game through reinforcement learning, and continuously improves its global decision-making capability and overall competitive level.

SenseCore Assists AI with “Graduation” in Five Weeks

As an infrastructure originally created by SenseTime, SenseCore provided powerful algorithm and computing power support for modeling, training, and verifyingDI-star.

Backed by SenseCore, DI-star completed approximately100 million matches in only five weeks, and finally reached the Grand Master level of human race.


DI-star used infrequent human tactics to win by surprise.

In addition, SenseCore also providedrich tools and technical components to help DI-star quickly try and error, experiment with a variety of new tactics, and drive the algorithm to iterate at an amazing speed in the reinforcement learning stage.

Undoubtedly, AI has made remarkable achievements in StarCraft 2. Nevertheless, it has not yet excelled human race in understanding this game. An obvious gap is that its performance fluctuates significantly when playing with characters of different races. Furthermore, AI still has much room for improvement in finding the optimal strategy for the game.

In the future, SenseTime will continue to rely on SenseCore, deepen the research and development of AI decision-making technologies, and improve the mutual competition, coordination, and cooperation capabilities of multiple agents in complex environments.

SenseTime Brings “AI Family Buckets” to the Game Industry

Boosted by both technological innovations and market demands, the game industry is rapidly stepping into the AI era.

For this reason, SenseTime established its Smart Decision and Game Business Department in 2021 to promote the integration, innovation, and development of AI and game industry.

Mr. Liu Yu, General manager of SenseTime Smart Decision and Game Business Department, said, “SenseTime will use its comprehensive AI technology capabilities, be centered around multiple aspects such asgame R&D,game operation,and game playing methods, andfully empower the development of the game industry.

For game research and development,SenseTime’s intelligent game content generation solutions can provide developers with AI technology based human motion capture and scenario modeling, which deliver natural and realistic effects.SenseTime’sintelligent game art generation solutions can intelligently generate various 2D and 3D game characters, as well as hand-painted and other scenario styles, helping developers quickly build different game scenarios and character shapes.

For game operation, SenseTime’s innovativecommunity robotandcommunity management SaaS helpdesk can provide intelligent customer service, cross-group invitation, group task and group score achievement management, high-value community information mining, high-value community member mining and other services, improve players’ activity and loyalty of players, and enhance users’ stickiness.

For game playing method innovations,SenseTime’s global game robots can provide a complete set of solutions from intelligent decision-making, intelligent service, intelligent operation and other intelligent agents inside and outside the game for different types of games, such as MOBA, SLG, RTS, FPS, MMO and RAC, so that AI has a more comprehensive multi-agent collaborative combat capability, and even simulates different players’ personalities, making itself become excellent teammates or opponents of human players.

With the continuous improvement of game AI decision-making intelligence, it is hopeful to redefine the relationship between players and AI, and bringa brand new game experience of human-machine symbiosis.

On the other hand, the in-depth application of AI technology in the field of game content generation will also help game developers reshape their game production process and significantly improve their creation efficiency.

In the future, SenseTime will take SenseCore as the cornerstone, promote the diversified application of AI in the game field through industry-leading full-stack capabilities, and work with industry partners to jointly promote the innovation and development of the game industry.