Simulation · no model and no key
Build an agent, then watch it play
Seven parts. With all of them on, the agent picks up a key and goes around a wall to the door. Turn one part off and see exactly where the task breaks.
- 1World
A grid, a wall, a key and a door. That's everything the agent can touch.
- 2Goal
Pick up the key, then stand on the door. Without a goal there is no end.
- 3State
Where it stands, and whether the key is in hand. Only an action changes that.
- 4Sensors
Without a sensor the policy doesn't know where the key is.
- 5Actions
Step or pick up. With pickup off, the key stays on the floor.
- 6Policy
A plan finds a way around. Walking straight gets stuck at the wall.
- 7Loop
Step, observe, decide, until success or until the budget runs out.
Plans a detour
Walks straight
The agent is at the starting point. Goal: pick up the key and reach the door. Versus: The agent is at the starting point. Goal: pick up the key and reach the door.
Keyboard: left and right arrows step, space plays. 1 of 10 · Start