This is a simple environment, created for the use with OpenAi gym and baselines. It features an n dimensional state and action space. The PointGoalEnv also features a goal space which has the same dimensionality. All spaces are continuous.
The state of the environment is a point. The action is a move command for the point and the goal is a point with a margin. The reward function is binary and gives a reward, only if the goal was reached.
The environment was created with the help of these tutorials [1, 2].
After cloning, the package can be installed with.
pip install -e .
There are automatic tests to verify the installation.
pip install pytest
pytest
If an agent uses the environment it should import gym_point. The package is compatible to python3.
The following environments are available:
The point environment has the standard 2 dimensional state and action space. The action is a move command to the state. The goal is a two dimensional circle with a position and a fixed margin. The goal position is set randomly with the initialization.
The environment is compatible with DDPG. Convergence is not yet proven.
python3 -m baselines.ddpg.main --env-id PointEnv-v0
This environment is an environment with variable goals. It implements gym.GoalEnv with all the required functions.
The environment is compatible with HER. Convergence is not yet proven.
python3 -m baselines.her.experiment.train --env PointGoalEnv-v0 --num_cpu 3