The code contains implementations of two algorithms
- ACRO - learns agent centric representations with a multi-step inverse dynamics model.
- InfoGating - adds an InfoGating bottleneck over a contrastive variant of the multi-step inverse dynamics model.
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.
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