PickApp

Package to perform apple picking experiments with a UR5 robotic manipulator and a 3-finger underactuated gripper.

Contact

Alejandro Velasquez
velasale@oregonstate.edu
Robotics PhD Student

Description

This package takes datasets of apple picks performed with a robotic manipulator UR5e and a robotic gripper as end effector. The picks are performed in a real tree and in an apple proxy that emulates the physics of the apple tree. Hence, this package is useful for the user to:

• Post-process data in a suitable form for Machine Learning classifiers.
• Compare time-series plots from real and proxy picks.
• Obtain statistics about the experiments, such as the distribution of cartesian and angular noise.
• Do basic machine learning to check if - with the given data - it is possible to train a classifier and predict the outcome of the pick (e.g. successful, failed).

For brief usage instructions, type:

$ python pickapp_compare.py --help
$ python pickapp_learning.py --help
$ python pickapp_stats.py --help
$ python pickapp_data.py --help

Examples

module: pickapp_compare.py

The following example analyzes variable 'Force_x', among the 'failed' picks case, and does the Dynamic Time Warping (DTW) analysis during the 'pick' phase.

$ python pickapp_compare.py --variable force_z --case failed --phase pick

It outputs a time-series plot with the closest real and proxy picks. It also outputs a .csv file with a list of the real and proxy picks that are comparable. Comparable picks are the ones where the pose of the robot with respect to the apple is the same. These files are stored in the sub-folder results.

module: pickapp_learning.py

The following example runs a Random Forest Classifier (RFC), with 10 experiments to account for the classifier's stochasticity, with a depth of 5 branches, and utilizes 5 features.

$ python pickapp_learning.py --experiments 10 --depth 5 --feature 5 --classifier rfc 

It outputs a boxplot with the classifier's accuracies during the experiments. The boxplot gets stored in a .pdf file, along with a .txt file with the confusion matrix of the best accuracy. These files are stored in the sub-folder results.

module: pickapp_stats.py

In the following example, the statistic analysis is run for the dataset 3_proxy_winter22_x1.

$ python pickapp_stats.py --dataset 3_proxy_winter22_x1

It outputs .pdfs with box-plots of the angular and cartesian noise. It also outputs a .txt file with Mean, SD and percentiles of each noise. These files are stored in the sub-folder results.

x_noises
Mean: -0.012SD: 0.014Percentiles: [-0.02  -0.01   0.003]
y_noises
Mean: -0.005SD: 0.01Percentiles: [-0.015 -0.007  0.   ]
z_noises
Mean: -0.012SD: 0.006Percentiles: [-0.015 -0.01  -0.005]
roll_noises
Mean: -4.846SD: 12.74Percentiles: [-10.  -5.   5.]
pitch_noises
Mean: -4.382SD: 12.627Percentiles: [-10.   0.   5.]
yaw_noises
Mean: 0.0SD: 0.0Percentiles: [0. 0. 0.]

Installation

Dependencies

I recommend installing the following packages:

• System related: os, sys
• User Interface related: tqdm, argparse
• Data related: csv, pandas, collections, ast
• Plots related: matplotlib, seaborn
• Math/Stat related: numpy, scipy, dtw
• Machine Learning related: sklearn

Data

Download the data subfolder, which includes all the features, metadata, and data from real and proxy domains picks

Choice 1: Install from PyPI

pip install pickApp

Choice 2: Download

Download the repo, and run the following code

python setup.py install

Tips

Tips to write better functions
https://pybit.es/articles/writing-better-functions/

Git cheat-sheet
https://education.github.com/git-cheat-sheet-education.pdf

MarkDown cheat-sheet
https://www.markdownguide.org/cheat-sheet/

Packaging Python Projects
https://packaging.python.org/en/latest/tutorials/packaging-projects/