rlai-lab/relod

An efficient remote-onboard architecture for real-time Reinforcement Learning

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): Youtube video

Supported Algorithms

• Soft Actor Critic (SAC)
• Proximal Policy Optimization (PPO)

N.B: All vision-based experiments use Random Augmented Data (RAD) to improve sample efficiency

Supported Tasks

UR-Reacher	Franka-VisualReacher

Create-Reacher	Vector-ChargerDetector

Choice of hyper-parameters for UR5 experiments

Hyper-parameter	Value
Replay buffer	100K
Actor step size	3e-4
Critic step size	3e-4
Entropy coefficient step size	3e-4
Batch size	256
Discount factor	0.99
Update every $k$ steps	2
Num. update epochs every $k^{th}$ step	1
Actor MLP hidden sizes	[512 512]
Critic MLP hidden sizes	[512 512]
Warm-up time steps	1000
Adam optimizer betas	[0.9, 0.999]
Initial temperature	0.1
Neural network activation	ReLU

Installation instructions

1. Download Mujoco and license files to ~/.mujoco
2. Install miniconda or anaconda
3. Create a virtual environment:

conda create --name myenv python=3.6    # Python 3.6 is necessary
conda activate myenv

3. 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

4. Install packages with:

pip install -r requirements.txt
pip install .

Run experiment

 python task_ur5_visual_reacher.py  --work_dir "./results" --mode 'l' --seed 0 --env_steps 200100 

Franka-VisualReacher

The code for the Franka task can be found in this branch.

Cite

Wang, Y.⋆, Vasan, G.⋆, & Mahmood, A. R. (2023). Real-time reinforcement learning for vision-based robotics utilizing local and remote computers. In Proceedings of the 2023 International Conference on Robotics and Automation (ICRA).