s-tian / bench-press Goto Github PK

This repository contains code for operating the tactile sensor testbench (modified 3 axis CNC), as well as model training code written in PyTorch.

It accompanies the paper OmniTact: A Multi-Directional High Resolution Touch Sensor, by Akhil Padmanabha, Frederik Ebert, Stephen Tian, Roberto Calandra, Chelsea Finn, and Sergey Levine, to appear in the International Conference on Robotics and Automation (ICRA) 2020.

Paper website: https://sites.google.com/berkeley.edu/omnitact

To run a control experiment on a dummy environment or on the testbench, create a configuration file and run it per the following example:

python run/run.py <experiment_config> <num_trajectories>

For a concrete example, the command python run/run.py keyboard_control_dummy 5 rolls out 5 trajectories in a dummy environment, using a policy which queries the user to enter actions via the keyboard.

To train a model using data collected using the run script, create a model configuration file and train the model using

python scripts/train_policy.py <model_config>

for example

python scripts/train_policy.py models/experiments/y_reg/y_reg_rand.yaml

trains a model which regresses the XYZ location of the end effector from tactile images.

Further hardware, electronics, firmware, and software documentation is provided in /doc/README.md.