This repository contains classical control benchmarks for evaluating robustnesses of control and reinforcement learning algorithms. It can be used as zero-shot control performance evaluations. It is built upon OpenAI Gym.

Clone repository, then 'pip install -e .' or 'pip3 install -e .' based on your environment.

Or you can manually install dependencies:

- numpy
- gym

You can run our test example by:

For pendulum,

python unstable_pendulum.py

For cartpole(continuous action),

python unstable_cartpole_cont.py

It's an inverted pendulum in gym environment. The sample results of the two different winds are shown below:

Sine wave side wind	Random side wind

It's a cartpole (continuous action) environment. The sample results of the two different winds are shown below:

Sine wave side wind	Random side wind

Simply import environments from 'unstable_gym'. For examples, for inverted pendulum:

from unstable_gym.unstable_pendulum import UnstablePendulumEnv
env = UnstablePendulumEnv(wind_type="sine", max_wind=1.0)

obs = env.reset()
for step in range(500):
    action = env.action_space.sample()
    nobs, reward, done, info = env.step(action)
    env.render()

For cartpole:

from unstable_gym.unstable_cartpole_cont import UnstableCartPoleContEnv
env = UnstableCartPoleContEnv(wind_type="sine", max_wind=1.0)

for ep in range(10):
    obs = env.reset()
    for step in range(1000):
        action = env.action_space.sample()
        nobs, reward, done, info = env.step(action)
        env.render()
        if done:
            break

There are two options for "wind type":

1. "sine" : sine wave side wind
2. "random" : random side wind

You can also adjust the magnitude of the side wind (in [N]): "max_wind"

You can test the robustness of MPPI and Smooth_MPPI