Whatizit?
Kinann is a Javascript library for building, training and using artificial neural networks (ANNs) that model real world kinematic errors.
Overview
Kinematic models for robots rarely match their implementations--axes may be slightly misallgned, part dimensions may differ from nominal, etc. To handle real world conditions, kinematic models are often extended to include error parameters. Unfortunately, the resulting kinematic models are often cumbersome and unwieldy to work with.
Kinann lets you create an artificial neural network (ANN) that bridges the gap between a simple, ideal kinematic model and any given implementation of that kinematic model. Kinann will handle all the error corrections automatically after proper calibration and training. Kinaan doesn't actually need to know the precise kinematics of your model--all it does is model the mismatch between ideal and actual coordinates. As long as your robot is precise, Kinann will make sure that your robot moves accurately to application coordinates:
IdealKinematics + Kinann = CalibratedRobot
Kinann kinematic error regression ANNs can be linear or even polynomial. Linear Kinann networks are often sufficient for Cartesian kinematics. However, you will need polynomial Kinann networks to deal with non-linear kinematics. For example, rotary delta kinematic errors often manifest as "bowl-shaped Z-plane errors".
Kinann is designed for building kinematic neural networks for Javascript robot applications. Although there are many neural network frameworks (e.g., synaptic.js, Tensorflow, etc.) that can be used to solve the kinematic error challenge, Kinann is optimized for kinematic modeling. Kinann should not be used to recognize cat pictures on YouTube.
Installation
Use npm to install kinann.
npm install kinann
See Also
- Kinann wiki...
- mathjs many thanks to MathJS for expression parsing and derivatives!