Methods for Processing and Plotting Tactile Shell Velociroach Data

###Install Dependencies Install virtualenv, if necessary. Virtual Environments allow you to create a sandbox where you can install only project-specific dependencies.

pip install virtualenv

Create a virtualenv, commonly called venv.

virtualenv venv

Activate the virtualenv created.

source venv/bin/activate

Thats it! Now you are ready to install project dependencies.

Install dependencies

pip install -r requirements.txt

###Install Tensorflow You have probably seen in the previous step that tensorflow might not have been installed properly using requirements.txt. If that is the case you will have to install it on your own. Here is a quick way to do it:

Activate venv if you haven't already

source venv/bin/activate

Install latest Tensorflow using pip

pip install --upgrade https://storage.googleapis.com/tensorflow/mac/tensorflow-0.7.1-cp27-none-any.whl

Go to this site to test your installation

jupyter notebook

utils.process_data_files(data_file, calibration_file) takes in a telemetry data_file (.txt) and a calibration file (e.g. calibration/out/C.mat) and returns a Pandas DataFrame, a Python dictionary-like object, with variables:

['time', 'Right Leg Pos', 'Left Leg Pos', 'Commanded Right Leg Pos', 'Commanded Left Leg Pos', 'DCR', 'DCL', 'GyroX', 'GyroY', 'GyroZ', 'AX', 'AY', 'AZ', 'RBEMF', 'LBEMF', 'VBatt', 'S1', 'S2', 'S3', 'S4', 'S5', 'S6', 'S7', 'S8', 'PowerR', 'PowerL', 'Energy', 'TorqueR', 'TorqueL', 'VMotorR', 'VMotorL', 'AngleZ', 'Fx', 'Fy', 'Fz', 'Mx', 'My', 'Mz', 'F_mag', 'A_mag', 'M_mag', 'Gyro_mag']

These variables are accessible in the following fashion.

import utils
DATA_FILE = "input/sliding9.txt"
CALIBRATION_FILE = "calibration/out/C.mat"
df = utils.process_data_files(DATA_FILE, CALIBRATION_FILE)
df['Fx']

Print header

import utils
DATA_FILE = "input/sliding9.txt"
utils.print_header(DATA_FILE)

Plot columns, display it, and save to out/basic.png

import utils
DATA_FILE = "input/sliding9.txt"
CALIBRATION_FILE = "calibration/out/C.mat"
df = utils.process_data_files(DATA_FILE, CALIBRATION_FILE)
utils.plot_columns(df, [["TorqueL", "TorqueR"], ["Left Leg Pos", "Right Leg Pos"], ["RBEMF", "LBEMF"], ["VMotorR", "VMotorL"], ["PowerR", "PowerL"], "VBatt", "AngleZ"], display=True, save_figure=True, output_dir="out/", output_filename="basic.png")

See this iPython Notebook for example usage.

####featurizer.py

Featurizes the data. Currently, we implement a 512 sample window (~.512s) with 50% overlap. Given that the leg frequency for our sliding experiments is 4Hz, our window would capture 1-2 complete leg cycles of data. For each window, we compute the following statistical metrics: [max, min, mean, std, skew, kurtosis] and frequency domain metrics: [entropy, energy] and also pairwise correlation.

You can view an iPython notebook showing our frequency domain features of an example telemetry data here.

####segment_data.py

Segments each data file into multiple segments and assign diff labels to each segment.

Usage:

python segment_data.py --input {filepath or directory} \
--output_dir {directory}

Example:

python segment_data.py \
--input experiment_data/yellow_roach/drag_experiments \
--output_dir terrain_identification

To segment:

1. Select each segment by clicking its beginning point followed by its end point. Each segment must be sequential and cannot overlap. There should be an even number of points selected, 2 for each segment. If there is an error while segmenting, such as an odd number of points selected, you have the option of redoing the segmenting, by entering [y]es on the error prompt.

2. After an even number of sequential points are selected, you have the option of viewing the chosen segments on the graph by pressing [d], redoing the segmentation procedure by pressing [r], quitting via [q], and keeping the points via [y]. If you choose to keep the points, you will be prompted to assign labels to each segment in order, starting from segment 0.

####classify.py

Classifies the data. Currently, five models are implemented: Random Forests, Gradient Boosted Trees, RBF SVM, a 2 hidden layer Neural Network with 100 nodes in each hidden layer, and an ensemble of the four previous classifiers. All of their average 10-fold cross validation accuracy are ~94-96% with the ensemble being the least variant at ~96%. Random forests and gradient boosted trees also give feature importances.

You can see the results here.

We've also performed predictions on a test example and the models performed remarkably well, especially the ensemble. The results are here.