Manipulation-Oriented Object Perception in Clutter through Affordance Coordinate Frames

Created by Xiaotong Chen, Kaizhi Zheng, Zhen Zeng, Shreshtha Basu, James Cooney, Jana Pavlasek and Odest Chadwicke Jenkins from Laboratory For Progress, the University of Michigan.
This is a pytorch implementation of the paper: Paper link, Video link.

Requirement

The code has been tested with

• CUDA 10.1
• Python 3.6
• PyTorch==1.1 & Torchvision==0.3

Dataset & Checkpoints

The synthetic dataset (containing 20k images, 53.9G) : link. You can download the files and store them under data/.

Our synthetic dataset is generated by NDDS, which is a plugin for Unreal Engine 4. You can find the usage in this repo.

You can download the models used for training and test at here.

The following checkpoints can be downloaded at here:

• endpoints_attention.pth: Attention Module + Endpoints head
• scatters_attention.pth: Attention Module + Scatters head
• norm_attention.pth: Attention Module + Norm vector head

You can store these checkpoints under weights/.

Running deep network for perception

Training

# Train a new (endpoints) model from pretrained COCO weight
python3 train.py

Test

# Load the checkpoint and the default output is shown in output/.
python3 test.py --resume endpoints_attention

Running robot experiment in CoppeliaSim platform

Download CoppeliaSim simulator at here (tested in version 4.1.0). The simulator is free of 3rd libraries, to run it

cd PATH_TO_CoppeliaSim
./coppeliaSim.sh

Open the simulator/sim_env.ttt scene in File/Open Scene..., you will be able to see a tabletop scene setting.

Run robot experiment after setting appropriate parameters in simulator/main.py

python3 simulator/main.py

If you want to use random texture for objects, we provide some sample textures here. Extract and put this folder to path simulator/texture. Feel free to add or delete textures in the folder.

The robot manipulation code has two parts, one is in simulator/robot.py, including API calls to Coppeliasim, interface to perception network, calculation of end-effector poses for grasping and manipulation, and utility functions. Another is in Coppeliasim scene file remoteApiCommandServer file and written in lua, including API calls to manipulator IK, motion planning, scene object setup, grasping, etc.

For more debug information related to robot simulator, refer to CoppeliaSim official website and forum.