Jul 20, 2020

OpenSim Webinar - Inverse Kinematics: A Bayesian Versus Least-Squares Approach

Learn the basics of Bayesian inference and how it can be used for inverse kinematics

Updated Information (September 9, 2021)

The main result discussed in this webinar – that Bayesian inference performs better than least-squares for inverse kinematics -- has been determined to hold true only for overly informative prior distributions.

You can read more about the follow-up analysis in “Comparing the Performance of Bayesian and Least-Squares Approaches for Inverse Kinematics Problems” by Pohl, et al.

Thank you to our presenter, Todd Pataky, for alerting us to this update. The corrected source code, along with the corrigendum article pre-print and extended explanations, are available at https://github.com/0todd0000/BayesIK

Recording

There is no recording, please refer to the updated information above.

Details

Title: Inverse Kinematics: A Bayesian Versus Least-Squares Approach
Speaker: Todd Pataky, Kyoto University

Abstract

When estimating joint kinematics from a set of noisy marker measurements, most inverse kinematics (IK) approaches aim to minimize errors by distributing them amongst the markers in a least-squares sense. This talk will describe how Bayesian calculations can be used to maximize the probability that a specific set of joint angles would produce the observed marker positions. This computation of marker positions from joint angles is referred to as a forward kinematics model or forward model. Preliminary Bayesian IK results show order-of-magnitude improvement over least-squares estimates, with Bayesian IK outperforming least-squares IK in over 90% of a large number of random simulations involving planar rotations, with an approximately tenfold decrease in rotation estimate error. This talk will assume no knowledge of Bayesian statistics, and will work from fundamental Bayesian concepts to simple IK models, then finally to explain why Bayesian IK appears to outperform least-squares approaches.

You can learn more about the work presented in the webinar by reading the publications:

• Bayesian inverse kinematics vs. least-squares inverse kinematics in estimates of planar postures and rotations in the absence of soft tissue artifact (2019)
• Bayesian vs. least-squares inverse kinematics: Simulation experiments with models of 3D rigid body motion and 2D models including soft-tissue artefacts (2020)

Supplementary material and code snippets from the 2019 article are also available.