contact_state_classification

This is the repo for the robotics project "contact state classification using machine learning".

Requirement

• numpy
• pandas
• tslearn (for Shapelets classifier)
• visdom
• facets
• tensorflow

How to use

1. Download the dataset from https://drive.google.com/drive/folders/1GyiogHXgIxUiuVfkc2BFoljfxg-VJTo3 and put it in contact_state_classification/tests/1908/hfv/csd_result/
2. Check contact_state_classification/config.py, Add the features you want to use to SIMPLE_FEATURES and COMPLEX_FEATURES according to the table blow, SIMPLE_FEATURES being features with only one dimension and COMPLEX_FEATURES being features with more than one dimension.
3. Then check the other parameters. If CIRCULAR_SPLICING is on, the results of the local exploration are repeated once, to achieve a circular filling effect. INTERPOLATION_METHOD is used to control the method used for interpolation. UPSAMPLING_RATE is used to control the density of interpolation. N_SPLITS is used to control the size of the Stratified KFold partition.
4. Run main.py, and check result in the console or write to logfile using cs_classifier::log_to_csv
5. Visualize the result using plotter.py, you need to turn on the visdom server before running this script.

Resources

Google Drive: https://drive.google.com/drive/folders/1GyiogHXgIxUiuVfkc2BFoljfxg-VJTo3 Datasets e.g., RoboticsProject2510.pkl and demo video e.g., wiggling_example.mp4 are saved here.

Feature Value Explanation

Feature	Description
label	label for contact state
lock_type	different type of latch lock
init_q	initial joint state
init_base_ht_ee	initial end-effector(ee) pose in the Cartesian space on the robot base frame
init_image	None
init_rs_image	initial image on the robot end-effector from a realsense(rs) camera
init_usb_image	initial image from an external perspective from a usb camera
desired control values
d_ee_velocity	desired ee velocity on the ee frame
d_ee_theta	desired ee velocity but in the spherical coordinate
d_ee_phi	desired ee velocity but in the spherical coordinate
observed state values
dist	observed moved distance after execution an exploration action (currently 12 actions for one exploration behavior in total)
base_s_ee	observed moved distance as form of directed vector after execution an exploration action
error_q	joint value difference. Once the robot finished a exploration action, it should move back to the initial joint pose (init_q). This feature shows how good the robot moved back to the initial pose
obs_ee_theta	observed ee velocity in the spherical coordinate after sending the d_ee_velocity to the robot
obs_ee_phi	observed ee velocity in the spherical coordinate after sending the d_ee_velocity to the robot
init_ee_ft_robot	observed force/torque sensor reading on the ee frame before controlling one exploration action
curt_ee_ft_robot	observed force/torque sensor reading on the ee frame after controlling one exploration action

Background Knowledge

• Panda Robot Frame Information: https://frankaemika.github.io/docs/control_parameters.html

  Frame 0 = Base Frame

  Frame F = End-Effector Frame

• Spherical Coordinate: https://upload.wikimedia.org/wikipedia/commons/d/dc/3D_Spherical_2.svg

  We transform the translation (x, y, z) into spherical coordinate to reduce the dimensionality from 3(x, y, z) to 2 (theta, phi). As we are only interested in moving direction namely theta and phi