willwhitney/Sectar

Self-Consistent Trajectory Autoencoder: Hierarchical Reinforcement Learning with Trajectory Embeddings

SeCTAr

Code for Self-Consistent Trajectory Autoencoder: Hierarchical Reinforcement Learning with Trajectory Embeddings

Running

Experiment scripts are in exps/, environments are located under traj2vec/envs, main algorithms are in traj2vec/algos/vaepdentropy.py and traj2vec/algos/vae_bc.py

Installation

• Download

  • sectar
  • rllab
  • doodad
  • baselines

• Add these repos to your python path and follow instructions for setting them up.

• Install Mujoco instructions here

• Modify traj2vec/launchers/config.py to point to the appropiate paths.

• Create the conda env sectar with

conda env create -f environment.yml

Logging

The log dir for the scripts is set to data. You can plot recorded results by giving the exp log dir to traj2vec/viskit/frontend.py.