The Ledger · Entry 004 · Games
Atari 2600 games from raw pixels
Ceded to machinesSince 2015Criterion performanceLast reviewed 2026-08-17
Event
DeepMind's DQN, learning from pixels and score alone, reaches at least 75 percent of a professional human tester's score on 29 of 49 Atari games; in 2020, Agent57 clears the human baseline on all 57 games.
The benchmark is a fixed suite of emulated games with a score signal, and the human baseline is one professional tester after about two hours of practice per game. Agent57 closed the last hard cases (Montezuma's Revenge, Pitfall, Skiing), at the cost of billions of frames of play per game, far more experience than any human. Sample efficiency and transfer to new games remain open; raw score on this suite is settled.