Skip to content
/ TD3 Public
forked from sfujim/TD3

Benchmarking TD3 and DDPG on PyBullet

Notifications You must be signed in to change notification settings

georgesung/TD3

 
 

Repository files navigation

Benchmarking TD3 and DDPG on PyBullet

This repo contains benchmark results for TD3 and DDPG using the PyBullet reinforcement learning environments:

  • Ant
  • HalfCheetah
  • Hopper
  • InvertedPendulum
  • InvertedDoublePendulum
  • Reacher
  • Walker2D

In the TD3 paper by Fujimoto et al., results were reported for the MuJoCo version of these environments. PyBullet is a free and open-source alternative to MuJoCo, with no license fees and no hardware lock (MuJoCo personal licenses are limited to 3 physical machines, which means you cannot run simulations in the cloud, e.g. AWS/GCP/etc).

This repo itself is a fork of the official TD3 code from the original authors of TD3, Fujimoto et al. All results presented here are derived using the original authors' algorithm implementations. Thus, references to "TD3", "OurDDPG", and "DDPG" below refer to Fujimoto et al's implementations.

Per the original repo, results were generated with Python 2.7 and PyTorch 0.4. Specific to this repo, PyBullet 2.4.8 was used. To obtain these results, run:

./run_experiments.sh

Learning curves

Below are the learning curves for TD3, OurDDPG, and DDPG, using the algorithms in 'TD3.py', 'OurDDPG.py', and 'DDPG.py'. Results for "TD3" and "OurDDPG" should serve as the PyBullet counterpart to the MuJoCo-based results in the original paper. Note the results for "DDPG" do not correspond to those presented in the TD3 paper (see original README file), but still serve as a good reference point nonetheless.

Per the TD3 paper, in the learning curves, the solid line represents the average over 10 trials, whereas the shaded area represents half a standard deivation over those 10 trials. The curves were smoothed uniformly for visual clarity, using scipy.ndimage.uniform_filter(data, size=7) (the window size was chosen arbitrarily by me).

HalfCheetahBulletEnv HopperBulletEnv Walker2DBulletEnv AntBulletEnv ReacherBulletEnv InvertedPendulumBulletEnv InvertedDoublePendulumBulletEnv

For more details on how the curves were generated, see 'plot_results.py', and also the original TD3 paper. To generate the plots yourself:

python plot_results.py

Pre-trained models

Pre-trained models can be downloaded from here.

Agent visualization

Here is an example of how to visualize a saved agent in action:

python eval.py --policy TD3 --env_name HalfCheetahBulletEnv-v0 --filename TD3_HalfCheetahBulletEnv-v0_0 --visualize

For more details, please refer to the code in 'eval.py'.

About

Benchmarking TD3 and DDPG on PyBullet

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 96.2%
  • Shell 3.8%