UNREAL / A3C / A3C-LSTM agent for MINOS

This is an implementation of the UNREAL agent described in Sim-Real Joint Reinforcement Transfer for 3D Indoor Navigation by Zhu et al. 2019 for use with the MINOS environment multimodal indoor simulator. This implementation is directly adaptated from https://github.com/miyosuda/unreal with a thin gym wrapper around the MINOS framework. The original implementation README.md file is here. This code also implements the A3C and A3C-LSTM baseline agents (as ablations of the full UNREAL agent -- see options.py for relevant flags).

Installation

Follow OUR MINOS installation instructions. Confirm that the minos package is available by running python3 -c 'import minos; print(minos)'.

Training baselines

Checkout git branch minos/train and unreal/master

Start experiments using invocations such as the following:

python3 main.py --env_type indoor --env_name roomgoal_suncg_sf

Refer to options.py for available arguments that control the hyperparameters and agent architecture.

Adversarial Transfer

Refer to adda

Policy Mimic

Checkout git branch unreal/mimic for policy mimic.

Set paths in code and run:

python3 main.py --env_type indoor --env_name roomgoal_mp3d_s

Testing

Checkout git branch unreal/test for testing.

python3 main.py --env_type indoor --env_name roomgoal_mp3d_sf