Orbit

Orbit is a open-source project and a collection of reinforcement learning environments. If you want to contribute to this project by creating your own environment then please drop a mail to shivajbd@gmail.com. You can know how to create an environment through this blog:

Requirements

• python - 3.7
• keras - 2.4.3
• tensorflow - 2.2.0

Checkout random_policy.py to see how to use an environment.

Orbit contains following RL environments as of today.

Environment 1: Paddle

Author: Shiva Verma

The task is to take the ball on paddle. Inbuilt turtle library is used to create the environment.

Action space (3)

• 0 - move paddle to left
• 1 - do nothing
• 2 - move paddle to right

State space (5)

• x position of paddle
• x and y position of ball
• x and y velocity of ball

Reward function

• +3.0 - when paddle hit the ball
• -3.0 - when ball touchs the ground
• -0.1 - when paddle moves

Episode termination

• Episode ends when ball touchs the ground.

Environment 2: Jump

Author: Vinod Kumar

The task to dodge the kangaroo from moving balls. Inbuilt turtle library is used to create the environment.

Action space (2)

• 0 - jump
• 1 - do nothing

State space (13)

• y position of kangaroo
• x and y position of all moving ball

Reward function

• +5.0 - when kangaroo dodge a ball
• -1.0 - for each jump
• +0.1 - for each time step

Episode termination

• Episode ends when kangaroo touchs a ball.

Environment 3: Cannon

Author: Sathish Kumar

The task to hit the dropping balls. pygame library is used to create the environment.

Action space (4)

• 0 - rotate left
• 1 - rotate right
• 2 - shoot bullet
• 3 - do nothing

State space (8)

• x position of cannon
• x and y position of all balls
• sin of angle of cannon

Reward function

• +5.0 - hitting the ball
• -0.5 - shooting bullet
• -0.1 - rotating canon
• -5.0 - episode termination

Episode termination

• Episode ends when a ball touchs the line.