learning-forces

Reference code for the paper What Does the Person Feel? Learning to Infer Applied Forces During Robot-Assisted Dressing.

Compiling on Linux?

./build_linux.sh

Test out force map estimation

You can try out our pretrained models by running:

cd physx_simulation/pysim/
python gownforcemaps.py
python shortsforcemaps.py

Full training process

Record 5,400 dressing sequences for both dressing tasks. This takes several hours on Amazon EC2:

python autogown.py 150
python autoshorts.py 150

Edit the last two lines in datapreprocess.py:
Replace 2016-08-31_arm and 2016-09-06_leg with the respective directories created from data collection above.

Run python datapreprocess.py

Edit the last two lines in learning.py so that the linked directories match the new ones created by datapreprocess.py.

Run python learning.py

Edit the second to last lines in both gownforcemaps.py and shortsforcemaps.py to link to the new directories created by learning.py.

Run python gownforcemaps.py or python shortsforcemaps.py to test out the newly trained LSTMs for force map estimation.

Quantitative Results

variationtestvizmse.py will generate all of the plots presented in the paper.

The test set data for these plots can be regenerated via:

python variationtestgown.py
python variationtestshorts.py

Setting up on Amazon EC2?

Checkout this GitHub repository for setup scripts and necessary dependency installations.

Prerequisites

python 2.7
Keras
SciPy
Tensorflow or Theano
plotly
PyOpenGL - 3.1.0 or greater
PyOpenGL_accelerate
pybind11
HDF5
h5py
DART
Eigen (C++ library)