Agents that Listen: High-Throughput Reinforcement Learning with Multiple Sensory Systems

Train an agent to play VizDoom with multi sensory inputs. Trained using sample factory

Paper: https://arxiv.org/abs/2107.02195

Website: https://sites.google.com/view/sound-rl

Tested on Ubuntu 18.04 64-bit.

• Install miniconda for Python 3.7: https://docs.conda.io/en/latest/miniconda.html

• Clone the repo: git clone https://github.com/hegde95/Agents_that_Listen.git

• Create and activate conda env:

cd Agents_that_Listen
conda env create -f environment.yml
cd ..
conda activate sound

This will install sample-factory into the environment too

• Clone the required ViZDoom repo that contains the sound state space: git clone https://github.com/hegde95/ViZDoom_with_Sound.git

• Build and install the new environemnt:

cd ViZDoom_with_Sound
python setup.py build && python setup.py install
cd ..

• Enter the ViZDoom_Sound folder and run the following command to start a training run on the instruction scenario discussed in the paper

cd Agents_that_Listen
python -m rl.train --algo=APPO --env=doomsound_instruction --experiment=doom_instruction --encoder_custom=vizdoomSoundFFT --train_for_env_steps=500000000 --num_workers=24 --num_envs_per_worker=20

• To view a rollout of the learned policy run the following line:

python -m rl.enjoy --algo=APPO --env=doomsound_instruction --experiment=doom_instruction

• To record a demo of the learned policy run the following command:

python -m demo.enjoy_appo --algo=APPO --env=doomsound_instruction --experiment=doom_instruction

Citation

@inproceedings{hegde2021agents,
  title={Agents that Listen: High-Throughput Reinforcement Learning with Multiple Sensory Systems},
  author={Hegde, Shashank and Kanervisto, Anssi and Petrenko, Aleksei},
  booktitle={To Appear in IEEE COG 2021},
  year={2021}
}