Semi-Supervised Haptic Material Recognition using GANs
Z. Erickson, S. Chernova, and C. C. Kemp, "Semi-Supervised Haptic Material Recognition for Robots using Generative Adversarial Networks", 1st Annual Conference on Robot Learning (CoRL 2017), 2017.
Project webpage: http://healthcare-robotics.com/mr-gan
Download the MREO dataset
Compact dataset (1 GB) (can be used to compute tables 1, 2, 3, 4, and 6): https://goo.gl/WiqSjJ
Full processed dataset (20 GB) (can be used to compute all tables in paper): https://goo.gl/FnXfgM
Raw data collected on the PR2 (10 GB): https://goo.gl/DNqPib
Dataset details can be found on the project webpage.
Running the code
Our generative adversarial network is implemented in Keras and includes the feature matching technique presented by Salimans et al.
GAN results presented in tables 1, 3, and 6 can be recomputed using the command below (requires compact dataset). This takes several hours with a GPU.
python mr_gan.py --tables 1 3 6
Neural network and SVM results from tables 2 and 4 can be recomputed using the commands below (requires compact dataset).
python mr_nn.py --tables 2 4 python mr_svm.py --tables 2 4
Recompute results presented in table 5 (requires full dataset).
python mr_gan.py --tables 5
Generate plots. This requires plotly.
Collect new data with a PR2.
rosrun fingertip_pressure sensor_info.py & python contactmicpublisher.py & python temperaturepublisher.py & python collectdataPoke.py -n fabric_khakishorts -s 100 -w 0.1 -l 0.1 -ht 0.06 -v python collectdataPoke.py -n plastic_fullwaterbottle -s 100 -l 0.03 -ht 0.08