The goal is to train the agents 'Ant-v3' and Humanoid-v3' from the MuJoCo environments of OpenAI Gym to move on the circumference of a circle centered on the origin while remaining within a safety region smaller than the radius of the circle.

The Environments were inspired by the circle environments in https://arxiv.org/pdf/1705.10528.pdf and are directly modified from the implementations in OpenAI Gym.

Go to environment directory

cd mujoco-circle

then install by

pip install -e .

After installation is complete, you can create an instance of the environment with gym.make('mujoco_circle:AntCircle-v0') (For HumanoidCircle, replace 'AntCircle-v0' with 'HumanoidCircle-v0').

The code snippet below generates an episode of maximum length 1000 using a random policy from the AntCircle environment.

import gym

env = gym.make('mujoco_circle:AntCircle-v0')
state = env.reset()
for _ in range(1000):
    action = env.action_space.sample()
    next_state, reward, done, info = env.step(action)
    if done:
        break
env.close()