Maze

This is a maze environment based on Minigrid. The overall task is partially observable navigation, and the agent must move around in a maze to reach the goal.

Action Space

7 dimensional discrete action space, where only the first three are used. We do not mask out invalid actions.

Observation Space

The size of the observation depends on the agent's view size (k), which is by default 5. The observations is kxkx3, consisting of floats, indicating which tile is in each of the cells around the agent.

Reward Function

By default, penalize_time is True, which means the reward upon reaching the goal is 1 - 0.9 * (t / T), where t is teh current timestep, and T is the maximum number of timesteps in the episode. If penalize_time is False, the reward is 1 upon reaching the goal and zero otherwise.

Level Generation

The function make_level_generator in jaxued.environments.maze.util returns a callable that generates a maze given an rng.

Rendering

We also provide JITted rendering functionality in jaxued.environments.maze.renderer. This can be used as follows:

env = Maze(max_height=13, max_width=13, agent_view_size=5, normalize_obs=True)
sample_random_level = make_level_generator(env.max_height, env.max_width, 25)
env_renderer = MazeRenderer(env, tile_size=8)
obs, state = env.reset_to_level(rng, sample_random_level(rng), env_params)

image = env_renderer.render_state(state)
plt.imshow(image)