qgallouedec / drl-grasping

Deep Reinforcement Learning for Grasping

Baselines and panda-gym

Code to train an agent on panda-gym environments.

Installation

Tested on Ubunutu 18.04 LTS.

1. install some dependencies

   sudo apt-get update && sudo apt-get install cmake libopenmpi-dev python3-dev zlib1g-dev

2. clone the repository

   git clone https://github.com/qgallouedec/drl_grasping

3. create a virtual environment, activate it and upgrade pip

   cd drl_grasping
   python3 -m venv env
   source env/bin/activate
   python -m pip install --upgrade pip

4. install dependecies

   pip install -r requirements.txt

Usage

Train

To train PandaPickAndPlace-v1 with seed 0 for 500000 timesteps, run

mpirun -np 8 python train.py PandaPickAndPlace-v1 0 500000

The learning is distributed over 8 MPI workers. For the moment, this number should not be modified.

Play

To play the learned policy, run

python play.py PandaPickAndPlace-v1

Process the results

Turn the brut results into .dat file, containing timesteps, median and quartiles.

python results_to_dat.py

It process all the training done so far.

Author

Quentin GALLOUÉDEC