We also demonstrate its effectiveness in generating robotic tasks.

We would like to thank all reviewers for reading our paper and providing constructive comments. Sometimes, the primary interest is to understand agent behaviors, and hence only the learning mode is needed. Alternatively, when all game inputs are known, the focus is on the intervention mode. In the final version, we will (i) explain in 2.1 how these We agree that it is neither rigorous nor necessary to assert that "most" Our work is inspired by the current interests on complex optimization-based layers. It is the first to treat VIs as individual layers in the end-to-end framework.

We also demonstrate its effectiveness in generating robotic tasks.

When game designer and entrepreneur Henk Rogers first encountered Tetris at the 1988 Las Vegas Consumer Electronics Show, he immediately knew it was special. "It was just the perfect game," he recalls. "It looked so simple, so rudimentary, but I wanted to play it again and again and again … There was no other game demo that ever did that to me." Rogers is now co-owner of the Tetris Company, which manages and licenses the Tetris brand. Over the past 30 years, he has become almost as famous as the game itself. The escapades surrounding his deal to buy its distribution rights from Russian agency Elektronorgtechnica (Elorg) were dramatised in an Apple TV film starring Taron Egerton.

A little more than a decade ago, the video-game designer Davey Wreden experienced a crippling success. In October, 2013, he and a collaborator, William Pugh, released the Stanley Parable HD, a polished and expanded version of a prototype that Wreden had developed in college, and which he had made available, free of charge, two years before. Wreden and Pugh hoped that they might sell fifty thousand or so copies of the new version in the course of its lifetime. They sold that many on the first day. Wreden was twenty-five years old, and he had everything he'd ever wanted: money, success, recognition.

Ahead of the 21st Bafta games awards this April, the institution is running a public survey asking people to nominate the most influential video game of all time. As the survey points out, this is an open-ended question: early, groundbreaking titles such as Space Invaders and Pong regularly crop up as answers because they helped write the rules of the form, but on a personal level, the right game at the right time can be exceptionally influential, too. For players, it's often the games that made us feel differently about what games could do that feel the most influential. For a game designer, a film director, a writer or a musician, one particular game might inspire a whole creative era. Inspired by Bafta's survey, we asked people from across games and culture for their most influential game – and not one name cropped up twice.

In 2005 I received a copy of World of Warcraft for my birthday. The game clocked in at 3 gigabytes--a behemoth by the standards of the early 2000s, so big that it had to be distributed across four different CDs. I installed those discs onto our creaking, overworked family PC and, hours later, created my first avatar: a humble dwarf paladin named Pumaras, who set off to explore a realm he would soon call home. World of Warcraft was a singular experience, and completely unlike the lonesome corridors of Halo or Call of Duty. Millions of living, breathing human beings logged on to the game at the same time.

Minecraft remains remarkably popular a decade or so after it was first released, thanks to a unique mix of quirky gameplay and open world building possibilities. A knock-off called Oasis, released last month, captures much of the original game's flavor with a remarkable and weird twist. The entire game is generated not by a game engine and hand-coded rules, but by an AI model that dreams up each frame. Oasis was built by an Israeli AI startup called Decart in collaboration with Etched, a company that designs custom silicon, to demonstrate the potential of hardware optimized to power transformer-based AI algorithms. Oasis uses a transformer AI model, similar to the one that powers a large language model--only trained, apparently, on endless examples of people playing Minecraft, to dream up each new video frame in response to the previous one and to user input like clicks or mouse moves.

In 2004, Holly Longdale was a game designer on EverQuest, then the champion of a new genre of video game that allowed for multiplayer role-playing on a huge scale. In these online fantasy worlds, players could quest together rather than alone, adding a fascinating new social – and competitive – dimension to the static, offline role-playing that Holly's generation had grown up with. But whenever she could, Longdale would sneak in a few hours playing EverQuest's main competitor instead. That game was World of Warcraft (WoW). "There were so many moments in WoW I was envious of," she says, "and completely lost in. I remember running through Ashenvale as a Night Elf Hunter and the music and the ambience – there was a mood you couldn't deny. Then I saw another player running in the opposite direction, a Druid who buffed me on their way by. That was when I knew I was going to be in this for the long-haul."

Generating simulations to train intelligent agents in game-playing and robotics from natural language input, from user input or task documentation, remains an open-ended challenge. Existing approaches focus on parts of this challenge, such as generating reward functions or task hyperparameters. Unlike previous work, we introduce FACTORSIM that generates full simulations in code from language input that can be used to train agents. Exploiting the structural modularity specific to coded simulations, we propose to use a factored partially observable Markov decision process representation that allows us to reduce context dependence during each step of the generation. For evaluation, we introduce a generative simulation benchmark that assesses the generated simulation code's accuracy and effectiveness in facilitating zero-shot transfers in reinforcement learning settings. We show that FACTORSIM outperforms existing methods in generating simulations regarding prompt alignment (e.g., accuracy), zero-shot transfer abilities, and human evaluation. We also demonstrate its effectiveness in generating robotic tasks.