Requirements

• Ubuntu (18.04 or latest LTS release)
• Up-to-date Nvidia driver
• 64 GB ram (~ 6 gb per parallel environment)

Introduction

This repository showcases how to make an "intermediate" level project with iGibson, it includes:

• Distributed training with ray
• declarative configuration of experiments with hydra
• using a custom model override for ray capable of mixed observation modalities
• rendering the environment with a parallelized evaluation environment during training
• implementing a custom task, custom termination condition, and custom reward function
• saving/loading agent state
• pinning all dependencies, and seeding the environment, for reproducible RL

Installation

Deterministic installation via conda-lock:

git clone [email protected]:mjlbach/ig_navigation
conda install -c conda-forge 'conda-lock[pip_support]'
conda-lock install -n ig_navigation
conda activate ig_navigation
pip install -e .

Regular installation (reproducible result not guaranteed)

git clone [email protected]:mjlbach/ig_navigation
conda env create -f environment.yml
conda activate ig_navigation
pip install -e .

Running

Before running, you will want to specify the path to ig_assets and ig_dataset:

export GIBSON_ASSETS_PATH=/home/michael/Documents/ig_data/assets
export IGIBSON_DATASET_PATH=/home/michael/Documents/ig_data/ig_dataset

This can be done automatically via a .envrc file and direnv.

To run with the default settings (experiments will be dumped to ray_results in the current folder:

python train.py ++experiment_save_path=$HOME/ray_results

This repo uses hydra to allow a mix of declarative and argument based configuration, select the experiment with +experiment=path:

python train.py +experiment=search ++experiment_name=my_experiment_name ++experiment_save_path=$(pwd)/ray_results

Note: This will take ~64 gb of ram. If this is not feasible, you can lower the number of environments used with an additional hydra flag, or by overriding in the experiment yaml (the default is at configs/experiments/search.yaml:

python train.py ++num_envs=4

Reproducibility notes

Developer dependencies are locked with conda-lock

conda-lock -c pytorch -c conda-forge -p linux-64

Running on 8 environments, the policy (standard PPO) should converge to a success rate of 95% on the single object visual navigation task in the empty Rs_int scene after about 300,000 environment steps, which takes approximately 1 hour on an RTX 3090/AMD 5900x processor.