Natural Language Directed Reinforcement Learning

Download Spatial-Reasoning (SPR) data

./download_spr_data.sh

Training

Training with vanilla VIN network on SPR with terminal state as input

python train_spr_planning.py

Training with joint representation model with VIN approach on SPR with Imitation Learning

Example kick-off training VIN model on SPR data as propsed in the report

python train_spr_global.py --epochs 30  --k 10 --save_dir trained_models/ --actions 4 --l_h 150 --map_type [local|global|joint]

Acknowledgements

Thanks to the following Authors for sharing data and code that served as a basis for our work

• https://github.com/JannerM/spatial-reasoning
• https://github.com/kentsommer/pytorch-value-iteration-networks