Source code of the paper In a Nutshell, the Human Asked for This: Latent Goals for Following Temporal Specifications @ ICLR 2022 and the early preprint version titled as "Relational Deep Reinforcement Learning and Latent Goals for Following Instructions in Temporal Logic" @ ICML 2021 Workshop on Uncertainty and Robustness in Deep Learning.
This repository requires Python3.6 with three libraries: numpy, tensorflow (V2.5), and gym.
python run_experiment.py --visual --pretrained_encoder --mapType=minecraft-insp --syntax=TTL2 --network=PrediNet --mem --multimodal --num_neurons=50 --seedB=0 --batch_size=512 --mem_type=rim
This runs the BRIM$^{LG}$ from the main document in the minecraft-inspired benchmark, (once a pretrain encoder is available), remove "--pretrained_encoder" if you want to train the agents in one go (note that it will take longer to converge, specially in the Minigrid setting and in the larger maps).
run_experiment.py contains instructions for each command