QDgym is a framework for Quality Diversity (QD) Optimization benchmarks. It is built on OpenAI Gym and uses the same interface. QDgym is meant to unify benchmarks for QD Optimization algorithms and provide a common interface that allows easy comparisons between different QD algorithms. QDgym was born out of the desire to not perpetuate the creation of new benchmarks that are difficult/impossible to use for other researchers when developing benchmark tasks for the PGA-MAP-Elites algorithm. These benchmarks were therefore implemented as this stand-alone open-source module with no proprietary software dependencies.

For those new to QD Optimization, here are three exellent introductory papers in no particular order:

• Quality Diversity: A New Frontier for Evolutionary Computation
• Confronting the Challenge of Quality Diversity
• Quality and Diversity Optimization: A Unifying Modular Framework

Also checkout: https://quality-diversity.github.io/

The current version is a v0.1 with a only the evaluation tasks of PGA-MAP-Elites implemented (plus humaniod). My plan is to add any new tasks I develop as part of my research under this format.

Currently five tasks are implemented that build on PyBullet robotic locomotion benchmarks where the task is to discover diverse ways to walk (Diversity aspect), where for each type of walk find the fastest walk (Quality aspect).

• Behaviour Descriptor (BD): Proportion of contact time each foot has with the ground over a walk. The BD in the QDAnt task is thus 4-D while QDHalfCheetah, QDHumanoid and QDWalker have 2-D BDs. The BD for QDHopper is 1-D.
• Fitness: Walking distance over 1000 simulation steps/walking speed.

• Observations: The robots current COG height, x, y and z velocity, roll, pitch and yaw angles and the relative position of the robots joints.
  • QDWalker2DBulletEnv-v0: 22-dimensions.
  • QDHalfCheetahBulletEnv-v0: 26-dimensions.
  • QDAntBulletEnv-v0: 28-dimensions.
  • QDHopperBulletEnv-v0: 15-dimensions.
  • QDHumanoidBulletEnv-v0: 44-dimantions
• Action: The torque to apply to each joint.
  • QDWalker2DBulletEnv-v0: 6-dimensions.
  • QDHalfCheetah: 6-dimensions.
  • QDAntBulletEnv-v0: 8-dimensions.
  • QDHopperBulletEnv-v0: 3-dimensions.
  • QDHumanoidBulletEnv-v0: 17-dimentions.
• Behavioural Descriptor: The time proportion each foot of the robot is in contact with the ground (scaled to range from 0 to 1).
  • QDWalker2DBulletEnv-v0: 2-dimensions.
  • QDHalfCheetahBulletEnv-v0: 2-dimensions.
  • QDAntBulletEnv-v0: 4-dimensions.
  • QDHopperBulletEnv-v0: 1-dimension.
  • QDHumanoidBulletEnv-v0: 2-dimensions.

Installation is easy with pip. Just do:

pip3 install git+https://github.com/ollenilsson19/QDgym.git#egg=QDgym

this will also automatically install the dependencies requred (gym, pybullet, numpy).

If the installation fails, it will likely be due to your pip version being too old for the requirements of opencv (gym dependency). See: https://pypi.org/project/opencv-python/. The solution is likely to upgrade pip.

pip install --upgrade pip

To run tasks with QDgym follow the below structure:

# imports
import gym
import QDgym

# make environment
env = gym.make("env_name")
# reset environment
state = env.reset()
done = False
# eval loop
while not done: 
    action = controller.select_action(state)
    state, reward, done, info = env.step(action) 
# get fitness and BD
fitness = env.tot_reward
behaviour_descriptor = env.desc # alternatively behaviour_descriptor = info["bc"] for compatability with the code released for MAP-Elites-ES
# close environment
env.close()

The above example assumes there is a defined controller with a method select_action that for a given input (state) outputs the desired action. Replace env_name with your desired task (QDWalker2DBulletEnv-v0, QDHalfCheetahBulletEnv-v0, QDAntBulletEnv-v0, QDHopperBulletEnv-v0, QDHumanoidBulletEnv-v0).

For more detailed example use see main.py and vectorized_env.py in the PGA-MAP-Elites implemetation.