https://github.com/takuseno/configurable-control-gym

Configurable control tasks based on default environments included in OpenAI Gym
https://github.com/takuseno/configurable-control-gym

Last synced: about 1 month ago
JSON representation

Configurable control tasks based on default environments included in OpenAI Gym

• Host: GitHub
• URL: https://github.com/takuseno/configurable-control-gym
• Owner: takuseno
• License: mit
• Created: 2020-03-31T04:53:28.000Z (over 4 years ago)
• Default Branch: master
• Last Pushed: 2020-05-07T05:55:09.000Z (over 4 years ago)
• Last Synced: 2024-10-23T04:07:00.595Z (2 months ago)
• Language: Python
• Homepage:
• Size: 22.5 KB
• Stars: 3
• Watchers: 4
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# configurable_control_gym

This repository provides configurable control tasks based on default environments included in [OpenAI Gym](https://github.com/openai/gym).

This work is inspired by [Packer et al, "Assessing Generalization in Deep Reinforcement Learning", 2018](https://arxiv.org/abs/1810.12282).

This program is especially useful to the following deep reinforcement learning research projects:

- assessing generalization performance

- domain randomization

- robust continuous control

## install

```

$ pip install git+https://github.com/takuseno/configurable-control-gym

```

## supported environments

All environments are adapted to continuous action-space even if the original environment supports only discrete action-space.

| name | original environment | configurable parameters (default value) |

|:-:|:-:|:-:|

| ConfigurableCartPole-v0 | CartPole-v0 | force(10.0), length(0.5), mass(0.1) |

| ConfigurablePendulum-v0 | Pendulum-v0 | force(10.0), length(1.0), mass(1.0) |

| ConfigurableAcrobot-v0 | Acrobot-v0 | length(1.0), mass(1.0), moi(1.0) |

| ConfigurableMountainCar-v0 | Continuous_MountainCar-v0 | force(0.001), mass(0.0025) |

| ConfigurableBipedalWalker-v0 | BipedalWalker-v2 | torque(80), friction(2.5), length(34) |

## usage

### set fixed parameters

```py

import gym

import configurable_control_gym

env = gym.make('ConfigurableCartPole-v0', length=2.0)

```

### set ranges of parameters

Multiple parameter ranges can be set at each parameter.

At every `reset` call, the parameters are uniformly sampled from the given ranges.

```py

import gym

import configurable_control_gym

env = gym.make('ConfigurableCartPole-v0', length=[[0.1, 2.0]],

               mass=[[0.001, 0.1], [0.2, 0.3]])

env.reset() # length is uniformly sampled from [0.1, 2.0]

            #, and mass is uniformly sampled from union of [0.001, 0.1] and [0.2, 0.3].

```