Turning a commercial game into an AI training environment
A game becomes a training environment when an agent can observe it, act in it, reset it and be scored, at scale and without a human at the screen. This guide covers the design choices that decide whether an environment is useful to a lab, and what to ask of a provider.
By adding the five things a game was never built to need: headless running on servers, time control, resets with fixed seeds, a defined observation and action space, and tasks with automatic scoring. There are two routes. A black box drives the normal build through pixels, keyboard and mouse — it needs only a licence and a build, and reaches a pilot in weeks. An engine integration exposes game state and native resets directly, giving exact determinism and faster-than-real-time play, at the cost of project access.
What does an environment need that a game does not have?
Games are built for one player at a time, in real time, with a screen. Training needs the opposite:
- Headless running on servers, many copies at once.
- Time control: faster than real time where possible, or stepped frame by frame.
- Resets and seeds: start any episode from a known state and repeat it exactly.
- An agent interface: a defined observation and action space.
- Tasks and rewards: goals with automatic scoring.
Pixels or game state?
Pixels with keyboard and mouse treat the agent like a human player. It works for every game, transfers to other games and suits world models and agents that learn from video. It is costlier to run and harder to learn from.
Structured game state (positions, inventories, resources) with discrete actions is cheaper and faster to learn from, and suits planning research and language-model agents. It requires engine access and a decision on what to expose.
Good providers offer both on the same game, so the lab can choose per experiment.
Black box or engine integration?
| Black box | Engine integration | |
|---|---|---|
| Needs from the studio | Licence and build | Licence and project access |
| Observation | Pixels | Pixels and game state |
| Speed | Real time | Faster than real time |
| Resets | Save games, VM snapshots | Native, instant |
| Determinism | Approximate | Exact with fixed seeds |
| Time to first pilot | Weeks | Weeks to months |
A practical path is to start a pilot as a black box, then move to engine integration once the lab and studio commit.
What makes an environment worth paying for?
- Realism: what the agent learns should transfer beyond the game.
- A wide state and action space: many situations, many valid strategies.
- Customisation: procedural or configurable maps, rules and conditions, so agents generalise instead of memorising.
- Good tasks: scoring that rewards the behaviour you want and is hard to game. Reward hacking is a known weakness of environments (TechCrunch: Silicon Valley bets big on environments to train AI agents).
- Accessibility: agents that work out of the box. If a lab has to build its own tooling to use the game, it churns.
Why isolation matters
A provider that runs environments for several labs can, in principle, observe how each lab's agents behave. Labs treat that as distillation risk and avoid shared providers. The answer is architectural: physically separate hardware per lab, no stored agent data, lab-held access keys and direct licensing between lab and studio. See how benned implements this.
Questions to ask an environment provider
- Is our environment on hardware no other customer uses?
- What do you store from our runs, and for how long?
- Who can access our environment, and is that logged?
- Do we contract with the rights holder directly?
- Which observation and action spaces are available, and how fast is a step?
- How are tasks scored, and can we define our own?
Running a pilot? benned sets up a licensed, isolated game environment in weeks.
Request a pilotFrequently asked questions
What is an RL environment?
A world an AI agent can observe and act in, with resets and automatic scoring, so the agent can learn by trial and error. Games are natural environments because they already have rules, goals and feedback.
Can any video game be used to train AI agents?
Technically almost any game can run as a black box with pixels and keyboard input. Legally you need a licence from the rights holder, and the most useful games are realistic, varied and customisable.
What is distillation risk in AI training?
The risk that a third party learns from how a lab's models behave and passes that knowledge on. Labs limit it by keeping training environments isolated and avoiding shared intermediaries.