ReLoD: The Remote-Local Distributed System for Real-time Reinforcement Learning on Vision-Based Robotics Tasks
ReLoD uses a wired local and a wireless remote computer to perform real-time learning, an appealing setting for industrial learning systems. It is a generalist RL system for learning with real robots from scratch! Check out how ReLoD learns to perform vision-based tasks on UR5 and Roomba (iRobot Create 2):
- Soft Actor Critic (SAC)
- Proximal Policy Optimization (PPO)
N.B: All vision-based experiments use Random Augmented Data (RAD) to improve sample efficiency
- UR5-VisualReacher
- Create-Reacher
- Vector-ChargerDetector
- Download Mujoco and license files to ~/.mujoco
- Install miniconda or anaconda
- Create a virtual environment:
conda create --name myenv python=3.6 # Python 3.6 is necessary
conda activate myenv
- Add the following to ~/.bashrc:
conda activate myenv
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/<username>/.mujoco/mjpro210/bin # Change based on mujoco version
export MUJOCO_GL="egl" # System specific
and run:
source ~/.bashrc
- Install packages with:
pip install -r requirements.txt
pip install .
- Wang, Y., Vasan, G., & Mahmood, A. R. (2022). Real-time reinforcement learning for vision-based robotics utilizing local and remote computers. arXiv preprint arXiv:2210.02317.