The goal is is to land a spacecraft in the moon in an specific location (the landing zone), without crashing.
Check ARLIE documentation and examples for detailed information on how to train policies in an environment in ARLIE.
The actions that can be taken on the spacecraft are the activation of the Yaw & Roll left/right, Pitch up/down, and the main engine boost. Of course no action can be taken as well.
The observation space is composed by the spacecraft position, rotation angle, speed, angular velocity (3D vectors), and the contact of each leg with the ground.
The default reward sums points inversely proportional to the distance to the landing zone and the velocity of the spacecraft.
Its score function penalize the usage of the engines (0.15 for the main one, 0.015 for the sides), and considers landing rough as a crash. The maximum points for landing (without taking into account previous steps) are 100, while the minimum for crashing far is -100.