arl-affpose-ros-node

I used arl-affpose-ros-node with the following repos:

1. LabelFusion for generating real images.
2. NDDS for generating synthetic images.
3. arl-affpose-dataset-utils a custom dataset that I generated.
4. pytorch-simple-affnet for predicting an object affordance labels.
5. densefusion for predicting an object 6-DoF pose.
6. barrett-wam-arm for robotic grasping experiments. Specifically barrett_tf_publisher and barrett_trac_ik.

ROS Architecture

Here is an overview of our Architecture.

Object Pose for Grasping

There are four main coordinate frames are used to grasp an object:

1. base link of the manipulator
2. camera frame
3. object frame
4. end effector frame

Note that the camera frame is a dynamic transform as we mounted our camera on our arm. See barrett_tf_publisher.

Note the object frame was determined either using marker-based methods, such as aruco_ros, or using deep learning, such as DOPE or DenseFusion.

Note that we used a 8-DoF Barrett Hand for the end effector frame. Which has +/- 17.5cm from tip to the center of the palm. Note that two-finger grippers require the object pose to be accurate within +/- 2cm.

Packages

• Ubuntu 18.04
• Cuda 10.0
• Python 2.7: 'conda create --name AFFDFROSNode python=2.7'
• Pytorch 1.4: 'conda install pytorch==1.4.0 torchvision==0.5.0 cudatoolkit=10.0 -c pytorch'

Requirements

$ conda env create -f environment.yml --name AFFDFROSNode