Skip to content

README.md

Latest commit

 

History

History
24 lines (16 loc) · 1.37 KB

README.md

File metadata and controls

24 lines (16 loc) · 1.37 KB
Raw

Usage

The code contains implementations of two algorithms

  1. ACRO - learns agent centric representations with a multi-step inverse dynamics model.
  2. InfoGating - adds an InfoGating bottleneck over a contrastive variant of the multi-step inverse dynamics model.

Installation

The provided requirements.txt file contains all main dependencies required to run this code (tested on Python 3.8). You will also need to download the offline datasets from v-d4rl paper or (depending on which kinds of distractors you want to test with) from the ACRO paper and have them in ./vd4rl path.

Running

Simply use the train.py script to run ACRO/InfoGating. Provide a task_name, offline_dir directory where the offline code is available and which algo to run. The dist_level argument is only used for naming result files.

Run ACRO

python train.py task_name=offline_cheetah_run_expert offline_dir=/path/to/dataset/vd4rl/main/cheetah_run/expert/ seed=1 algo=acro dist_level=none

Run InfoGating

python train.py task_name=offline_cheetah_run_expert offline_dir=/path/to/dataset/vd4rl/main/cheetah_run/expert/ seed=1 algo=infogating dist_level=none