Anticipation of Hand-object Contact Configuration - Montbonnot-Saint-Martin, France - Université Grenoble Alpes

Université Grenoble Alpes

Entreprise vérifiée

Montbonnot-Saint-Martin, France

il y a 1 semaine

Posté par:

Sophie Dupont

beBee Recruiter

Description

Anticipation of Hand-Object Contact Configuration for Object Manipulation:

Réf

-
ABG-117550

Sujet de Thèse 30/10/2023
Contrat doctoral
Université Grenoble Alpes
Lieu de travail
MONTBONNOT
Auvergne-RhôneAlpes
France
Intitulé du sujet
Anticipation of Hand-Object Contact Configuration for Object Manipulation
Champs scientifiques
Informatique
Robotique
Mots clés deep learning, trajectory prediction, manipulation, contact
Description du sujet:

Context and Motivation:

Learning hand-object contact configurations typically use large data sets consisting of 3D object shapes together with hand (robot gripper) poses, that together determine stable object-hand configurations.

Still most approaches require complicated refinement strategies that disregard infeasible hand-object configurations from physically plausible ones [1].

In this work package we propose to jointly model both, the manipulation strategy together with the final object-hand configuration that should be reached.

The intuition behind this approach is that knowledge about the appropriate action (e.g. manipulation strategy - sliding, picking) for a specific scenario determines the prediction of final object-hand configurations.

Thus unrealistic configurations that are either not reachable or interfere with the environment in a physically not plausible won't be even raised as a possibly feasible option.

Summary
The knowledge about possible manipulation strategies and the object lead to feasible hand-object configurations. This projects aims at determining feasible hand-object contact configurations for object manipulation.

Approach:

Many learning based approaches that consider only the object and the manipulator, generate hand object configurations that are either not reachable or interfere with the environment in an harmful manner.

For realisation, we will capitalise on the recent trend of action prediction using conditional variational autoencoders (CVAE) [2,3].

CVAEs are versatile deep generative that extend the standard VAE by conditioning with auxiliary properties - thus are suitable to develop visual representations conditioned on activity information.

We will follow-up on existing concepts that will be further developed:

Learning a (latent) representation conditioned on highlevel manipulation strategies.
Semisupervised learning for learning from just few (but correct) examples. This stands in contrast to learning from massive amounts of noisy data.

Goal:

Modèle fiche de poste 2020 Developing representations that jointly model visual information and dynamic action understanding - motion and contact.

These representations will allow fundamentally new ways of interpreting grasps as dynamic motion trajectories that are not only defined by pre-computed static hand-object configurations, further this new way of modelling hand-object configurations will allow for incorporating information about a suitable manipulation strategy.

Prise de fonction:

01/01/2024
Nature du financement:
Contrat doctoral

Précisions sur le financement:

Présentation établissement et labo d'accueil:

Université Grenoble Alpes

Collaboration between Inria team THOTH and the multidisciplinary institute in artificial intelligence (MIAI) in Grenoble

Etablissement délivrant le doctorat:

UNIVERSITE GRENOBLE ALPES
MSc degree in computer science or similar field
Experience in robotics, machine learning, computer vision, and/or control, industrial experience is a plus.
Excellent software engineering and programming skills in C++ and/or Python
Interest in interdisciplinary research in the context of the MIAI Grenoble Alpes Institute
Excellent English writing and communication skills 19/11/2023