frankroeder / dreamduck

World Models Experiments for Duckietown

World Model DuckieTown

This repository is an implementation of the WorldModelsExperiments combined with forks of gym-duckietown and baselines.

There are three gym environments provided.

1. DreamDuck-v0: default environment dreamduck/envs/env.py

2. DreamDuck-v1: world model representation dreamduck/envs/realenv.py

3. DreamDuck-v2: dream environment dreamduck/envs/rnnenv.py

Getting Started

Installation

1. Create a virtual environment with python3 -m venv venv and activate it with source venv/bin/activate

• If the module is not present run sudo apt-get install python3-venv, otherwise make use of instructions for your OS

2. Install dependencies pip install -r ./dreamduck/envs/requirements.txt

3. Install baselines fork pip install git+https://github.com/Bassstring/baselines

4. Install this module pip install -e .

Manual Control

All three environments can be controlled manually:

Default environment

• python dreamduck/envs/env.py
• Use the flag -h for all options

World Model Interpretation of the real Environment

• python dreamduck/envs/realenv.py
• Show real observation next to world model interpretation with --debug

Dreaming without real Environment

• python dreamduck/envs/rnnenv.py
• Make use of flag --temp to control uncertainty

Training

baselines (0.1.5)

The baselines module provide a straightforward way of training an agent with different algorithms and settings out of the box.

With the following commands an agent is trained in the dream and evaluated in the real environment interpreted by our world model.

• python -m baselines.run --alg=ppo2 --env=DreamDuck-v2 --num_timesteps=2e7 --network=mlp --num_env=2 --save_path=./models/dreamduck_rnnenv_ppo2 --log_path=train_rnnenv_logs

• python -m baselines.run --alg=ppo2 --env=DreamDuck-v1 --network=mlp --num_timesteps=0 --load_path=./models/dreamduck_rnnenv_ppo2 --play

For debugging purposes invoke baselines with following environmental variable:

• DEBUG_BASELINES=1 python -m baselines.run --alg=ppo2 ...

Running headless training with xvfb

• Install xvfb sudo apt install xvfb -y
• Run xvfb-run -a -s "-screen 0 1400x900x24 +extension RANDR" -- python -m baselines.run --alg=ppo2 --env=DreamDuck-v0 --num_timesteps=2e7 --network=cnn_lstm --num_env=8 --save_path=./models/dreamduck_cnn_lstm_ppo2 --log_path=train_logs

If there are issues follow this instruction.

Presentation to initial Results

https://docs.google.com/presentation/d/1wxcVQcTnhOC700dCKF-H1ftDOI84jh5EPQZgoMzgHvc/edit?usp=sharing

Authors

Frank Röder & Shahd Safarani