Gymnasium Mars Lander

Gymnasium environment for the Mars Lander CodinGame puzzles:

• Mars Lander - Episode 1
• Mars Lander - Episode 2
• Mars Lander - Episode 3
• Mars Lander - Optimization

Action Space	Box(-1, 1, (2,), float32)
Observation Space	Box(-1, 1, (12,), float32)
Import	gymnasium.make("gymnasium_mars_lander:gymnasium_mars_lander/MarsLander-v0")

This package is inspired by the article of Antoine Broyelle: Learning to Land on Mars with Reinforcement Learning.

Installation

To install gymnasium-mars-lander with pip, execute:

pip install gymnasium_mars_lander

From source:

git clone https://github.com/Quentin18/gymnasium-mars-lander
cd gymnasium-mars-lander/
pip install -e .

For running on CPU with extras:

pip install -e .[training,testing,quality] --extra-index-url https://download.pytorch.org/whl/cpu

Environment

Action Space

The action is a ndarray with two continuous variables:

• The rotation change between -15 and 15 degrees.
• The thrust change between -1 and 1.

The values are normalized between -1 and 1.

Observation Space

The observation is a ndarray of 12 continuous variables:

• The distances in six directions from the current position.
• The rower horizontal and vertical speed, angle and thrust.
• The horizontal and vertical distances to the middle of the landing area.

The values are normalized between -1 and 1.

The following figure shows the sensors used:

Rewards

The rewards are described by the following table:

Condition	Reward
The rover leaves the frame	-150
The rover runs out of fuel	-150
The rover crashes outside flat ground with incorrect angle and speed	-100
The rover crashes with correct angle and speed	-75
The rover crashes on flat ground	-50
The rover approaches the landing area	0.01
The rover lands successfully	200 + Amount of remaining propellant

Starting State

The starting state is generated by choosing a random CodinGame test case. When the eval_env argument is False, some random augmentations are applied to the test case. For each test case, there are five starting positions in increasing order of difficulty. The starting position can be set with the start argument.

The following figure shows the starting positions:

Episode End

The episode ends if either of the following happens:

1. Termination: The rower lands on the landing area or runs out of fuel or crashes.
2. Truncation: Episode length is greater than 2000.

Arguments

• episode: episode number between 1 and 3. The default value is 2.
• start: starting position between -1 and 4. The default value is -1.
• eval_env: if True, the random augmentations are disabled. The default value is False.
• sequential_maps: if True, the maps are generated sequentially. The default value is False.

import gymnasium as gym

gym.make(
    "gymnasium_mars_lander:gymnasium_mars_lander/MarsLander-v0",
    episode=2,
    start=-1,
    eval_env=False,
    sequential_maps=False,
)

Trained agents

There is one trained agent for each episode:

Path	Episode
rl-trained-agents/ppo/gymnasium_mars_lander-MarsLander-v0_1/best_model.zip	1
rl-trained-agents/ppo/gymnasium_mars_lander-MarsLander-v0_2/best_model.zip	2

Usage

You can use RL Baselines3 Zoo to train and evaluate agents:

pip install rl_zoo3

Train an Agent

The hyperparameters are defined in hyperparams/ppo.yml.

To train a PPO agent for the Mars Lander game, execute:

python -m rl_zoo3.train \
  --algo ppo \
  --env gymnasium_mars_lander/MarsLander-v0 \
  --tensorboard-log logs \
  --trained-agent rl-trained-agents/ppo/gymnasium_mars_lander-MarsLander-v0_2/best_model.zip \
  --n-timesteps 10000000 \
  --log-interval 100 \
  --eval-freq 10000 \
  --eval-episodes 100 \
  --seed 42 \
  --gym-packages gymnasium_mars_lander \
  --conf-file hyperparams/ppo.yml \
  --progress \
  --env-kwargs "episode:int(2)" "start:int(-1)" "sequential_maps:True"

Enjoy a Trained Agent

To see a trained agent in action on random test cases, execute:

python -m rl_zoo3.enjoy \
  --algo ppo \
  --env gymnasium_mars_lander/MarsLander-v0 \
  --n-timesteps 1000 \
  --deterministic \
  --seed 42 \
  --gym-packages gymnasium_mars_lander \
  --load-best \
  --progress \
  --env-kwargs "episode:int(2)" "start:int(-1)" "sequential_maps:True"

To see a trained agent in action on CodinGame test cases, execute:

python -m scripts.enjoy --path rl-trained-agents/ppo/gymnasium_mars_lander-MarsLander-v0_2/best_model.zip

To record videos of a trained agent in action on CodinGame test cases, execute:

python -m scripts.enjoy \
  --path rl-trained-agents/ppo/gymnasium_mars_lander-MarsLander-v0_2/best_model.zip \
  --record-video

Tests

To run tests, execute:

pytest

Citing

To cite the repository in publications:

@misc{gymnasium-mars-lander,
  author = {Quentin Deschamps},
  title = {Gymnasium Mars Lander},
  year = {2026},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/Quentin18/gymnasium-mars-lander}},
}

References

• Gymnasium
• RL Baselines3 Zoo
• Stable Baselines3
• Mars Lander with Reinforcement Learning
• CodinGame Forum - Mars Lander Puzzle discussion

Author

Quentin Deschamps