Kinematic Regression

This repository implements a Kinematic Regression approach to obtain configuration space data of a soft robot directly from pixels. The method follows roughly the following sequence:

1. From a demonstration video of a soft robot, CV techniques are used to get the cartesian space pose for each of the N marked cross sections.
2. Based on a PCS parametrization and assuming an initial N segmentation, obtain the strain data for each of these segments.
3. Iteratively join adjacent segments which have similar strain behaviour, based on the average strain-space distance between pairs of consecutive segments.
4. The configuration of each new merged segment is determined by performing a one-segment inverse kinematics on the distal ends of the merged segment.

Installation

• Clone this repository
• Install dependencies (see below)

How to Use

• PCSKinematicRegression_comparison.py is the main script. Run this file to do the Kinematic Regression.
• soft_manipulator_curve_fitting.py implements the function get_task_pose which is responsible to extract the task space pose from the video.
• segment_merging_algorithm.py contains the algorithm for merging the initial N segments.
• utils.py has auxiliary functions used across the above three files.
• tradeoff_plots.py generates the position and orientation errors as functions of the number of segments (i.e. the threshold) chosen

Dependencies

• numpy
• cv2
• scipy
• matplotlib