Body Shape-Guided Motion Retargeting to Reduce Effort on Human to Humanoid Landmark Placements

Chan-Soon Lim,Dong-Soo Kwon

doi:10.1109/access.2021.3065140

Abstract

Motion retargeting reduces the animator’s efforts to create robot motion by adapting human motion. However, it still requires a number of manual landmark placements to achieve satisfactory whole-body retargeted motion. Therefore, to reduce efforts on placing landmarks for corresponding minor body parts, this paper proposes a whole-body motion retargeting method for a general humanoid robot that considers the body shape similarity in addition to traditional landmark-based similarity. An additional strategy that matches the volumetric distribution of the body shape between a human and a robot is presented as guidance to handle redundancy from fewer landmarks and to force consistent outcomes from ambiguous landmark placements. The kinematically constrained Gaussian mixture model, originally used as a volumetric model-based human tracking method, is adapted and modified to manage both the shape and the landmarks in the proposed method. The shape and landmark similarity metrics are respectively introduced, and the overall similarity metric is defined by composing the sum of both metrics with weighting coefficients to control the balance between two policies by animators. Then the expectation-maximization based optimization is utilized to calculate target robot angles with human demonstration frame by frame. Experimental results validate the effectiveness of body shape matching, controllability through weighting coefficients, noise robustness on self-retargeting, and generality on applying different humanoid robots.

Highlights

W E expect humanoid robots to move like us, yet their relatively simple and rigid body makes it difficult for them to mimick a highly articulate human movement with a deformable skin
We focused on the volumetric distribution of a body shape that can represent the overall body pose as a skeleton model but without ambiguity, and can be directly obtained from recent shape-based motion data [20]–[22]
We adapted and modified the kinematically constrained Gaussian mixture model (KC-Gaussian mixture model (GMM)) [27], which was originally used in human tracking, for our retargeting frameworks to cover both the body shape and the landmarks

Summary

Introduction

W E expect humanoid robots to move like us, yet their relatively simple and rigid body makes it difficult for them to mimick a highly articulate human movement with a deformable skin. Simulating human-like motion in a humanoid robot is not trivial, several studies have discovered that it is expected by the user [4], [5], and that it can improve enjoyment [6], relationship [5], and human-robot collaboration [7]. Feature-based motion retargeting methods have the strength to adapt to various robots with relatively less effort. These approaches solve the embodiment mapping problem by maximizing motion similarity calculated from feature correspondences defined automatically or manually between the human and the robot [11]. Feature-based approaches compose mapping strategy from feature correspondence, which requires relatively less animator’s labors.

Objectives

Results

Conclusion