MarsExplorer

Description

MarsExplorer is an environment for evaluating agents on grid-based terrain exploration and coverage. Based on the MarsExplorer environment by Koutras et al., agents control a rover navigating a procedurally-generated 2D grid with obstacles, using a simulated LIDAR sensor to reveal unknown terrain. The goal is to explore as much of the map as possible while avoiding obstacles and staying within bounds.

Capabilities

• Spatial reasoning and navigation planning on ASCII grid maps
• Obstacle avoidance from partial observations
• Exploration strategy under step budgets
• Adapting to procedurally-generated terrain of varying difficulty

Compute Requirements

MarsExplorer does not require a sandbox. All game logic runs in-process with minimal compute.

License

ORLv1.

Tasks

There are 1,000 training tasks across three difficulty tiers:

• Small (334 tasks): 11x11 grid, 5 obstacles, 150 max steps, LIDAR range 4
• Medium (333 tasks): 21x21 grid, 12 obstacles, 400 max steps, LIDAR range 6
• Large (333 tasks): 41x41 grid, 30 obstacles, 1000 max steps, LIDAR range 8

Each task uses a fixed seed for reproducible map generation. The agent starts at position (0, 0) and must explore the grid by issuing directional move commands.

Reward Structure

This is a dense, verifiable reward environment matching the original MarsExplorer reward structure. No LLM graders are used.

Per-step reward (every move):

• new_explored_cells - movement_cost (movement_cost = 0.2)
• Exploring new terrain yields positive reward; revisiting explored areas costs 0.2

Terminal rewards:

• 95%+ explored (success): +400 bonus
• Collision with obstacle: -400 penalty, episode ends
• Out of bounds: -400 penalty, episode ends
• Max steps reached: episode ends with final step reward

Tools

• move(direction): Move the rover one cell in a cardinal direction ("up", "down", "left", "right"). Returns an ASCII map observation showing the current state of exploration, along with step count, position, and exploration percentage.

Map Representation

The agent receives a text-based ASCII grid after each move:

Step: 15/400 | Position: (5, 3) | Explored: 34.2% (151/441)

    01234567890123456789
 0  ..............??????
 1  ..............??????
 2  ......##......??????
 3  .....@.#......??????
 4  ..............??????
 ...

Legend: @ = rover, # = obstacle, . = explored, ? = unexplored

Other Environment Requirements

There are no further environment requirements. MarsExplorer works out of the box with the OpenReward endpoint without any secrets.

Safety

MarsExplorer is a grid navigation task with no safety concerns. The agent interacts only with an abstract grid world and cannot affect any real systems.

Citations

@article{koutras2021marsexplorer,
  title={MarsExplorer: Exploration of Unknown Terrains via Deep Reinforcement Learning and Procedurally Generated Environments},
  author={Koutras, Dimitrios I. and Kapoutsis, Athanasios Ch. and Amanatiadis, Angelos A. and Kosmatopoulos, Elias B.},
  journal={Electronics},
  volume={10},
  number={22},
  pages={2751},
  year={2021},
  publisher={MDPI},
  doi={10.3390/electronics10222751}
}