Datasets for Data-Driven Deep Reinforcement Learning with Pybullet environments. This work is intending to provide datasets for data-driven deep reinforcement learning with open-sourced bullet simulator, which encourages more people to join this community.

This repository is built on top of d4rl. However, currently, it is impossible to import d4rl without checking MuJoCo activation keys, which fails the program. Thus, d4rl_pybullet.offline_env is directly copied from d4rl repository.

$ pip install git+https://github.com/takuseno/d4rl-pybullet

The API is mostly identical to the original d4rl.

import gym
import d4rl_pybullet

# dataset will be automatically downloaded into ~/.d4rl/datasets
env = gym.make('hopper-bullet-mixed-v0')

# interaction with its environment
env.reset()
env.step(env.action_space.sample())

# access to the dataset
dataset = env.get_dataset()
dataset['observations'] # observation data in N x dim_observation
dataset['actions'] # action data in N x dim_action
dataset['rewards'] # reward data in N x 1
dataset['terminals'] # terminal flags in N x 1 

• random denotes datasets sampled with a randomly initialized policy.
• medium denotes datasets sampled with a medium-level policy.
• mixed denotes datasets collected during policy training.

You can train Soft Actor-Critic policy on your own machine.

# giving -g option to choose GPU device
$ ./scripts/train -e HopperBulletEnv-v0 -g 0 -n 1

You can collect datasets with the trained policy.

$ ./scripts/collect -e HopperBulletEnv-v0 -g -n 1

You can collect datasets with the random policy.

$ ./scripts/random_collect -e HopperBulletEnv-v0 -g -n 1

Any contributions will be welcomed!!

This repository is formatted with yapf. You can format the entire repository (excluding offline_env.py) as follows:

$ ./scripts/format

This work is supported by Information-technology Promotion Agency, Japan (IPA), Exploratory IT Human Resources Project (MITOU Program) in the fiscal year 2020.