PyTorch implementation of TD3 and DDPG for OpenAI gym tasks
Branch: master
Clone or download
Latest commit c717e75 Nov 30, 2018
Type Name Latest commit message Commit time
Failed to load latest commit information.
learning_curves Camera-ready cleanup Jun 9, 2018 Pytorch 0.4 Update Aug 10, 2018 Pytorch 0.4 Update Aug 10, 2018 Update Aug 10, 2018 cleanup + 1e6 max buffer Nov 30, 2018 bugfix for envs with max_action < 1 Sep 18, 2018 bug in script Mar 1, 2018 cleanup + 1e6 max buffer Nov 30, 2018

Addressing Function Approximation Error in Actor-Critic Methods

PyTorch implementation of Twin Delayed Deep Deterministic Policy Gradients (TD3). If you use our code or data please cite the paper.

Method is tested on MuJoCo continuous control tasks in OpenAI gym. Networks are trained using PyTorch 0.4 and Python 2.7.


The paper results can be reproduced exactly by running:


Experiments on single environments can be run by calling:

python2 --env HalfCheetah-v1

Hyper-parameters can be modified with different arguments to We include an implementation of DDPG ( for easy comparison of hyper-parameters with TD3, this is not the implementation of "Our DDPG" as used in the paper (see

Algorithms which TD3 compares against (PPO, TRPO, ACKTR, DDPG) can be found at OpenAI baselines repository.


Learning curves found in the paper are found under /learning_curves. Each learning curve are formatted as NumPy arrays of 201 evaluations (201,), where each evaluation corresponds to the average total reward from running the policy for 10 episodes with no exploration. The first evaluation is the randomly initialized policy network (unused in the paper). Evaluations are peformed every 5000 time steps, over a total of 1 million time steps.

Numerical results can be found in the paper, or from the learning curves. Video of the learned agent can be found here.