The classic trap in this domain is "passes the benchmark, unusable in the game". One, policies do not transfer: a policy learned under a fixed protocol like Atari 100k collapses when resolution changes, when a key mapping changes, or after one version update, because it learned pixel statistics rather than the rules of the game - and commercial patches change both. Two, evaluation protocols are detached from real play: Procgen, MineDojo and BALROG each measure one facet (generalisation, open world, language understanding) and none represents "can an agent complete a real stretch of gameplay for a player", so benchmark readings in this domain are defaulted to needs-reproduction. Three, memory and state management are the true bottleneck: a long playthrough requires the agent to remember what happened tens of minutes ago, where an item was left and what an NPC said, and current context windows plus retrieval mechanisms cannot hold persistent state at that scale, with the failure showing up as walking the same path repeatedly or forgetting a puzzle already solved. Four, cost spills over even in resettable environments: one full playthrough costs far more inference than one conversation, an overnight automated run can burn a substantial token budget, which means large-scale self-play at current prices is only viable for a handful of labs. Five, the sim-to-real gap: control policies learned in a game expose the physics engine's approximation error the moment they reach a robot, and that is the fundamental reason games and embodied stay in separate domains.
