DeepMind, Google's AI lab, has a history of showing off AI capabilities through games and tormenting human opponents in the process. In 2016, AlphaGo defeated world Go champion Lee Sedol. In 2019, AlphaStar built enough additional pylons to defeat the pros. StarCraft II Player (yes, that’s a thing) Grzegorz “MaNa” Komincz won 5-0. And in 2020, the Atari57 scored higher than the average human player across 57 games on the Atari 2600.
However, the lab's latest AI news is a little different. Rather than designing a model to master a single game, DeepMind worked with researchers at the University of British Columbia to develop an AI agent that can play many disparate games. It is called SIMA (sexpandable IStructurable metersuper world beGentleman), this project is competitive; cooperative Play as the AI operates according to human instructions.
But SIMA wasn't just created to help sleepy players level up or accumulate resources. Rather, the researchers hope that by better understanding how SIMA learns in these virtual playgrounds, they can make AI agents more cooperative and helpful in the real world.
Choose your own AI adventure
DeepMind worked with developer studios to train and test SIMA across nine different video games. These included a huge (if sometimes boring) space exploration game. no man's skyviking survival game valheimSF factory builder satisfaction,Furthermore goat simulator 3 — A game where you play as a goat and spread mayhem and violence across a fictional Southern California city.
This is a diverse portfolio of games, but these nine were not selected because DeepMind is concerned about a future where only AI can protect us from extraterrestrials and jetpacking goats. It's from. Rather, these games share similar characteristics to the real world in important ways. That is, it takes place in his 3D environment, which changes in real time, independent of the player's actions. It also features a variety of potential decisions and unlimited interactions beyond just winning.
To teach SIMA how to navigate these worlds, the researchers trained the AI to understand language-based instructions, starting with simple gameplay videos. In some videos, two players are playing the game together. One player will give instructions and the other player will follow them accordingly. In other videos, players played the game freely, and the researchers later annotated the videos with written instructions explaining the on-screen actions. In both cases, the goal was to capture how language connects to movement and action within the game.
To ensure that SIMA's skills apply to any 3D environment, not just the one in which it was trained, the researchers posed several additional design challenges. These include:
- The environment never slowed down, allowing SIMA to slowly decide on its next action. It needed to run at normal speed in-game.
- SIMA only had access to the images and text on the screen, and the same information was at the human player's disposal. We do not receive any privileged information from the game's source code.
- SIMA could only interact with the game using the equivalent of a keyboard and mouse.
Finally, SIMA's goals in the game needed to be delivered in real-time by human players using natural language instructions. The AI was not allowed to simply play the game over and over again until it achieved a specific predetermined goal. In other words, SIMA had to learn how to play cooperatively with a human partner, rather than through trial and error through thousands of game trials.
For example, if your latest in-game objective is valheim We were supposed to collect 10 pieces of wood, but the human user told SIMA to collect 10 pieces of rock instead. Instead of cutting down trees, SIMA should start mining nearby rocks.
“This is an important goal for AI in general,” the researchers note. [LLMs, like ChatGPT] Powerful systems have emerged that can gather knowledge about the world and generate plans, but they currently lack the ability to take action on our behalf. ”
Jack of all trades, future expert
SIMA was evaluated across a total of 1,485 unique tasks across nine skill categories. Skill categories are more generalized throughout the game (navigation, construction, object usage, etc.). On the other hand, the tasks themselves may be more specific to the game environment. For example, “Go to the ship,” “Build the generator,” and “Cut that potato.”
Overall, SIMA successfully understood language-based instructions. Its success rate varied by task and skill category, but it performed significantly better than other AI agents used in the evaluation.
For example, it outperformed AI trained in a single environment, with an average improvement of 67%. The researchers also trained some version of SIMA on all the datasets. exclude For one thing. When these versions of SIMA played an absentee game, they performed nearly as well as the single-environment AI. These results suggest that SIMA skills are transferable across different 3D environments.
The researchers also pitted SIMA against an AI agent trained without verbal input.in no man's skyFor example, SIMA had an average success rate of 34%. Other AIs averaged only 11%, a performance researchers describe as “adequate but purposeless.” These results further suggest that language is an important factor in SIMA performance.
“This study is the first to demonstrate that an agent can understand a wide range of game worlds and perform tasks within them, following natural language instructions, just as humans can,” the researchers said.
That being said, don't expect SIMA to immediately help you conquer a particularly difficult game.in no man's skydespite strict screening criteria, human players achieved a 60% success rate.
Similarly, because these games are complex, often combining dozens of potential interactions and hundreds of objects at any given moment, the researchers say the instructions “can be completed in about 10 seconds.” limited to things. It remains to be seen whether future iterations of SIMA will be able to manage the long-term, multi-layered planning of an average group of middle school students. Mine Craft Players manage every Saturday night.
The team published their findings in a technical report.
Think beyond the game
SIMA is a work in progress, and as the researchers look to the future, they hope to expand the game portfolio to include new 3D environments and larger data sets to scale up SIMA's capabilities. . Still, they think preliminary results are promising.
“Learning to play just one video game is a technical feat for an AI system, but learning to follow instructions in a variety of game settings unlocks an AI agent that is more useful in any environment. “Our research shows how the functionality of advanced AI models can be translated into useful real-world actions through language interfaces,” they write.
If we want AI to exist and operate all over the world, whether in cars or robots, virtual environments give researchers a safer and less wasteful way to test these systems.
If the AI crashes the car it is learning to drive in the virtual world, it can try again after learning its lessons. Crashing a car in the real world causes much more damage, even if it's not fatal. Once trained, the same language instructions can be used to direct the AI outside of the virtual setting. This is possible even if the AI is trained as a goat.
“Doing so makes SIMA an ideal platform for conducting cutting-edge research on securely grounded languages and pre-trained models in complex environments, thereby helping to address fundamental challenges in AGI. ”, the researchers conclude.
