Physically Plausible Dynamic Full-Body Grasp Synthesis for Human-Object Interactions


Loading...

Author / Producer

Date

2022-09

Publication Type

Master Thesis

ETH Bibliography

yes

Citations

Altmetric

Data

Abstract

Human motion synthesis is challenging, especially the generation of full-body human-object interactions, as it involves a diverse set of motor skills and objects. Recent methods have shown success in modelling grasping motions, but they either neglect the full-body constraint or only synthesise motions until obtaining a grasp of the object. This thesis addresses these shortcomings by incorporating the dynamics (physical forces) with a physical simulation of a humanoid with articulated fingers and an object. We model the task of approaching, grasping and lifting the object in the reinforcement learning setting. To that end, we extend two existing methods based on manual reward engineering and adversarial motion imitation. While the manual reward engineering approach only requires a grasp and target reference, the motion prior method relies on entire reference motion sequences. We empirically validate that our extension of the manual reward engineering method successfully solves the grasping and lifting task on the training data but exhibits unrealistic motions. Moreover, we demonstrate that adapting the adversarial motion prior method effectively increases the naturalism of full-body human-object grasp synthesis but can exhibit visual artefacts and has a lower task success rate. Further development of the motion prior method to mitigate these deficiencies promises an exciting future research direction.

Publication status

published

External links

Editor

Contributors

Examiner : Christen, Sammy
Examiner: Aksan, Emre
Examiner : Kocabas, Muhammed
Examiner: Hilliges, Otmar

Book title

Journal / series

Volume

Pages / Article No.

Publisher

ETH Zurich, Advanced Interactive Technologies

Event

Edition / version

Methods

Software

Geographic location

Date collected

Date created

Subject

Organisational unit

03979 - Hilliges, Otmar (ehemalig) / Hilliges, Otmar (former) check_circle

Notes

Funding

Related publications and datasets