The world

An agent must find the object named in an instruction. It sees part of the maze and chooses one action at a time.

These lessons explain the ideas through small experiments. The study compares how to spend the label budget.

A maze and an instruction

For “go to the grey box,” the agent must stand next to the box and face it. One attempt is an episode. It ends on success or after actions.

About this environment

The task is , with connected rooms on a -cell grid. Across generated worlds, the lab found instructions using six colors and three object types.

An expert follows a route to the target. Each action can become an example for the agent to learn from.

What the agent sees

The agent sees a 7×7 patch ahead, with walls blocking its view. Each cell has three codes: object, color, and state. The instruction and facing direction complete the input. These are symbolic codes, not camera images.

A full maze and the object, color, and state arrays of one local observation. — Left: the full maze, shown for us. Right: the object, color, and state codes the agent receives.

Inspect an episode

Move the slider through one expert episode. The highlighted cells show what the agent can see at each step.

The three actions

Three actions are available: . Turning changes where it faces; moving forward advances one clear cell. Doors are already open.

The same starting state followed by the result of turning left, turning right, or moving forward. — Each action starts from the same position.

Can random actions succeed?

A policy is a rule for choosing actions. Picking randomly succeeds in of these episodes. The maze is small and the time limit is generous. A learned policy should improve on this baseline.

The random policy starts in the same world as the expert above.

Why memory comes next

The agent must choose an action from a partial view. The next lesson shows why that view alone may not be enough.