Colour Grid World¶

colour_grid_world is a 9 x 9 single-agent tabular gridworld.

Shared Gridworld Conventions¶

This environment is a single-agent tabular gridworld with an explicit stochastic transition model. It exposes a full transition matrix via get_transition_matrix().

It uses the standard gridworld action convention:

• 0: move left
• 1: move right
• 2: move down
• 3: move up
• 4: stay in place

When slip is enabled, the intended action is taken with high probability and the remaining probability mass is spread uniformly over the other actions.

Environment Details¶

• start state 0,
• goal state 80,
• slip probability 0.1,
• one special blue state, one green state, and one purple state.

The environment uses:

• observation space Discrete(81),
• action space Discrete(5),
• labels {"start"}, {"goal"}, {"blue"}, {"green"}, and {"purple"},
• cost 1.0 on "blue" and 0.0 otherwise.

Reward is sparse:

• 1.0 when the agent reaches the goal state,
• 0.0 otherwise.

The episode terminates when the goal state is reached.

This is a small benchmark for experiments where the reward target and the safety-relevant state are intentionally different.