I found the current implementation of PPO on continuous action space is whether somewhat complicated or not stable.
And this is a clean and robust Pytorch implementation of PPO on continuous action space. Here is the result:

All the experiments are trained with same hyperparameters. Other RL algorithms by Pytorch can be found here.

gym==0.18.3
box2d==2.3.10
numpy==1.21.2
pytorch==1.8.1
tensorboard==2.5.0

run 'python main.py', where the default enviroment is Pendulum-v0.

run 'python main.py --write False --render True --Loadmodel True --ModelIdex 400'

If you want to train on different enviroments, just run 'python main.py --EnvIdex 0'.
The --EnvIdex can be set to be 0~5, where
'--EnvIdex 0' for 'BipedalWalker-v3'
'--EnvIdex 1' for 'BipedalWalkerHardcore-v3'
'--EnvIdex 2' for 'LunarLanderContinuous-v2'
'--EnvIdex 3' for 'Pendulum-v0'
'--EnvIdex 4' for 'Humanoid-v2'
'--EnvIdex 5' for 'HalfCheetah-v2'

P.S. if you want train on 'Humanoid-v2' or 'HalfCheetah-v2', you need to install MuJoCo first.

You can use the tensorboard to visualize the training curve. History training curve is saved at '\runs'

For more details of Hyperparameter Setting, please check 'main.py'

Proximal Policy Optimization Algorithms
Emergence of Locomotion Behaviours in Rich Environments