Video games have played an important role in the development of AI. Many early demonstrations of machine learning involved teaching computers to play games. Ultimately, Google Deepmind's learning of Game Starcraft 2 proved that the machines can compete with us in many areas where we were previously uncontroversial champions.
Currently, the game is being used as a testbed to explore some of the most exciting new realms in AI, including autonomous agents, real robots, and even AGI exploration.
At this year's Game Developer Meeting, Google's Deepmind AI division demonstrated research into what is called the Scalable Instructable Multiworld Agent (SIMA).
The idea is to show that machines can navigate and learn within the 3D world of video game environments. Then, using what you learn, you will navigate completely different worlds and tasks, all using your own rules, using the tools you can use to solve problems.
While it may sound like children's play, this study could dramatically affect the development of agent AI used in our work and personal lives. So let's see what that means and whether we can solve the ultimate AI challenge that allows humans to create machines that can adapt to any situation.
Virtual World
Video games provide a great environment for training your AI, as the variety of tasks and challenges are almost infinite. Importantly, players usually solve these challenges using a standard set of tools accessed via game controllers.
This corresponds well with how AI agents tackle the problem by selecting the tool they use from predefined selections.
The game world also provides a safe, observable, scalable environment in which the effects of subtle changes on variables and behavior can be investigated at almost real cost.
Deepmind's Simas was trained in nine different video game environments, filmed from popular games, including No Man's Sky, Valheim and Goat Simulator. Agents were given the ability to interact and control the game using natural language commands such as “Pick up keys” and “Move to a blue building.”
Among the outstanding findings, the study showed that agents were highly effective in transferable learning. I've tame what I learned in one game and used it to improve in another game.
This was backed up by the observation that agents trained them to play eight out of nine games. One game has improved performance in games that are not trained than specialized agents trained in one game.
This dynamic learning ability is important in the world where agents work with us, and helps us explore, interpret and understand troublesome real-world problems and situations.
But when it's common for robots to assist with physical and digital tasks, why not look a little further?
Physical Robots
The development of real-world robots that perform physical tasks has accelerated in collaboration with the evolution of AI over the past decade. However, it is generally used only in large companies due to the high cost of training due to the role of specialists.
Using virtual and video game environments can dramatically reduce this cost. The theory is that transferable learning allows physical robots to use their hands, arms, or any tool to tackle many physical challenges, even if they have not encountered before.
For example, robots that effectively learn how to use their hands to work in warehouses may also learn how to use them to build a home.
Before releasing ChatGpt, Openai demonstrated research in this field. Dactyl is a robotic hand trained in a virtual simulation environment and learned how to solve Rubik's Cube. This was one of the first demonstrations of the possibility of transferring skills learned in a virtual environment into tasks in the complex physical world.
Recently, NVIDIA has explicitly developed the ISAAC platform with the aim of training robots to “learn” how to perform real tasks within a virtual environment.
Today, physical AI-assisted robots work in warehouse roles, agriculture, healthcare, delivery, and many other jobs. However, in most cases, these robots still carry out particularly trained tasks at the enormous cost of companies with very deep pockets.
However, new models of “affordable” robots are on the horizon. This year, Tesla plans to manufacture thousands of Optimus robots, and allocate many of them to work in factories. Chinese robot developer Unitree has also recently announced a $16,000 humanoid robot that can turn its hands on many tasks.
With robot prices dropping and AI brains becoming more powerful by the day, walking, talking about humanoid robots can step into everyday reality faster than we think.
Towards Agi?
Almost 30 years ago, the car recorded its first major victory over humans, beating Gary Kasparov in chess. Few people would have predicted that there would be a computer that could beat the world champion in any game, not just one game.
This ability to “generalize” information by obtaining knowledge from one task and using it to solve completely different things is traditionally exclusive to humans, but it can be changing.
All of these are very interesting for those chasing the Holy Grail of AI Development, Artificial General Information (AGI).
The evidence that agents like Simas from Deepmind can transfer learning from one virtual gaming environment to another suggests that they may be developing some of the qualities needed for AGI. It shows that we are gradually building up capabilities that can be applied to solving future problems.
Google, together with Openai, Anthropic and Microsoft, states that AGI development is the ultimate goal and clearly the logical endpoint of the current focus on agent intelligence. Can I place another part of the puzzle in a video game?
