Skip to content

Latest commit

 

History

History
81 lines (65 loc) · 2.73 KB

readme.md

File metadata and controls

81 lines (65 loc) · 2.73 KB
Raw

Introduction

This repository is the control module of CRAVES. This repository contains hardware drivers, pose estimator, a PID-like controller, and a RL-based controller.

Hardware Requirement:

  • USB Camera (Logitech C920)
  • Robotic Arm (OWI-535)

Software Requirement:

  • Python>=2.7, Gym, CV2, Matplotlib, Numpy, Pytorch, Pyusb

It is easy to install craves_control, just run

git clone https://github.com/zfw1226/craves_control.git
cd craves_control
pip install -e .

Modify Permission in Linux

If your OS is Linux, you need modify your system to allow all users access to the arm via usb, just copy the file 42-usb-arm-permissions.rules to /etc/udev/rules.d, running as:

sudo cp 42-usb-arm-permissions.rules /etc/udev/rules.d/42-usb-arm-permissions.rules

Virtual Environment

If you want to train/evaluate a new RL agent in virtual environment, please install Gym-UnrealCV.

Usage

Running A Simple Demo

Firstly, please place the camera to a position viewing the arm. After that run the following command:

python craves_control/demo_control.py --kp --pose 0 0 0 0

The arm will move to an expected pose (0, 0, 0, 0) in a few seconds.

Reacher

The reacher aims to move the arm to make the grip reach a expected location. The policy network is trained by DDPG, a conventional RL algorithm for continuous action. Firstly, you can evaluate the policy network in virtual environment, running:

cd ddpg
python main.py --gpu-ids 0 --rescale --test --env UnrealArm-ContinuousPose-v1 --model-dir models/best.pt

After that, the arm will move to nine points sequentially, as: reach-virtual

If your hardware is ready, you can run the pose estimator and RL controller jointly:

python main.py --gpu-ids 0 --rescale --test --env RealArm --model-dir models/best.pt

After running, the arm will automatically move to an initial position at first, and then reach a set of points one by one, as: reach1

It is also robust to different viewpoint, as: reach2

Citation

If you found CRAVES useful, please consider citing:

@article{zuo2019craves,
  title={CRAVES: Controlling Robotic Arm with a Vision-based, Economic System},
  author={Zuo, Yiming and Qiu, Weichao and Xie, Lingxi and Zhong, Fangwei and Wang, Yizhou and Yuille, Alan L},
  journal={CVPR},
  year={2019}
}

Related Work