rl-toolkit 3.1.1

The RL-Toolkit: A toolkit for developing and comparing your reinforcement learning agents in various games (OpenAI Gym or Pybullet).

RL toolkit

Papers

• Soft Actor-Critic
• Generalized State-Dependent Exploration
• Reverb: A framework for experience replay

Setting up container

# Preview
docker pull markub3327/rl-toolkit:latest

# Stable
docker pull markub3327/rl-toolkit:2.0.2

Run

# Training container (learner)
docker run -it --rm markub3327/rl-toolkit python3 training.py [-h] -env ENV_NAME -s PATH_TO_MODEL_FOLDER [--wandb]

# Simulation container (agent)
docker run -it --rm markub3327/rl-toolkit python3 testing.py [-h] -env ENV_NAME -f PATH_TO_MODEL_FOLDER [--wandb]

Tested environments

Environment	Observation space	Observation bounds	Action space	Action bounds
BipedalWalkerHardcore-v3	(24, )	[-inf , inf]	(4, )	[-1.0 , 1.0]
Walker2DBulletEnv-v0	(22, )	[-inf , inf]	(6, )	[-1.0 , 1.0]
AntBulletEnv-v0	(28, )	[-inf , inf]	(8, )	[-1.0 , 1.0]
HalfCheetahBulletEnv-v0	(26, )	[-inf , inf]	(6, )	[-1.0 , 1.0]
HopperBulletEnv-v0	(15, )	[-inf , inf]	(3, )	[-1.0 , 1.0]
HumanoidBulletEnv-v0	(44, )	[-inf , inf]	(17, )	[-1.0 , 1.0]

Results

Summary

Return from game

Environment	gSDE	gSDE + Huber loss
BipedalWalkerHardcore-v3(2)	13 ± 18	-
Walker2DBulletEnv-v0(1)	2270 ± 28	2732 ± 96
AntBulletEnv-v0(1)	3106 ± 61	3460 ± 119
HalfCheetahBulletEnv-v0(1)	2945 ± 95	3003 ± 226
HopperBulletEnv-v0(1)	2515 ± 50	2555 ± 405
HumanoidBulletEnv-v0	-	** ± **

---

Frameworks: Tensorflow, Reverb, OpenAI Gym, PyBullet, WanDB, OpenCV
Languages: Python, Shell
Author: Martin Kubovčík

v3.0.7 (June 1, 2021)

Features 🔊

• Reverb
• updated kernel_initializer for last layers
• without clipping the mean
• setup.py (package is available on PyPI)
• split research process into agent, learner and tester roles

v2.0.2 (May 23, 2021)

Bug fixes 🛠️

• update Dockerfile
• update README.md
• formatted code by Black & Flake8

v2.0.1 (April 27, 2021)

Bug fixes 🛠️

• fix Critic model

v2.0.0 (April 22, 2021)

Features 🔊

• Add Huber loss
• In test mode, rendering to the video file
• Normalized observation by Min-max method
• Remove TD3 algorithm