Definition
Plain language
A score for how much of an environment an AI agent has actually explored, like a fog-of-war percentage.
As stated in the literature
Exploration Checkpoint Coverage — a deterministic metric defined as the fraction of pre-enumerated environment checkpoints (rooms, objects, valid actions) an agent touches during a no-task rollout.
Also called: Exploration Checkpoint Coverage
Why it matters: It separates 'can the agent solve tasks' from 'does the agent actually explore,' which is increasingly important for evaluating curious, open-ended behavior.
For example, before any task is even given, an agent is dropped into an environment and ECC measures what fraction of the rooms, objects, and valid actions it discovers on its own.