Missions

A first part of the MeSMRise project will focus on learning (multimodal) representations and graphs of interactions structured by actions. While random or naïve action policies can be used for that, directed manipulations should be more efficient for learning object representations.
The postdoc candidate will thus focus on learning action policies and address the main following questions: - How to learn in an unsupervised manner to select actions that will lead to better representations? We will consider the active object manipulation learning setting and explore the use of intrinsic drives derived from SSL losses to learn manipulation policies. - How to learn hierarchical policies for object manipulation using SSL losses as active learning drives? We will study the impact of having access to these different levels of actions in a hierarchical policy case - How to leverage learned graphs of anticipations to guide learning of efficient policies? We will also investigate how to best leverage the inference information provided by learned anticipations from other WP to further guide the learning of the agent. Indeed the abstract representations (e.g. sensorimotor primitives) and higher level inferences in non-Markovian environments can be used for optimal action planning (e.g. using informed heuristic search algorithm D*), structure augmentations (e.g., to bind several small rotations as larger scale manipulation of a single object). This can be used to select the best course of action, to either optimize exploitation (e.g. for distinguishability) or exploration (e.g. biaising curiosity mechanisms). Finally, the postdoc candidate will be expected to contribute to the coordination with other project tasks and partners.

Activités

- Propose different learning strategies and architectures
- Evaluate those strategies using a 3D robotics environment simulating 3D objects manipulation
- Write scientific articles
- Coordinate with other partners of the project and contribute to integration

Compétences

The ideal candidate would have a PhD in a relevant field and:
• Strong experience and publication record in Machine Learning, especially Deep Learning and Reinforcement Learning for object manipulation and perception
• Experience with Active Learning, Intrinsic Motivations and/or Self Supervised Learning are strongly desirable.
• Experience with 3D robotics simulators
• Ability to interact face-to-face and remotely with different members of the consortium;
• Autonomy and proactivity in research activities and reporting

Contexte de travail

This posdoctoral position is part of the MeSMRise (Multimodal deep SensoriMotor Representation learning) ANR project (https://projet.liris.cnrs.fr/mesmrise/index.html)
The MeSMRise project proposes to take inspiration from the way human babies learn to explore their environment through actions that shape their multimodal experience. Inspired by the sensorimotor contingencies (SMC) theory, the main objective of the project is to study how action can structure the multimodal representations, learned with self-supervised learning (SSL) methods. This will be applied to 3D objects, perceived by vision and point cloud, and manipulated in virtual environments.
This postdoc position takes place in the third workpackage of the project related to active learning, focusing on learning action policies that allow efficient learning of object representations.
The candidate will work at Institut Pascal, in Clermont-Ferrand vicinity, and will interact with other project partners in Lyon and Grenoble.

Le poste se situe dans un secteur relevant de la protection du potentiel scientifique et technique (PPST), et nécessite donc, conformément à la réglementation, que votre arrivée soit autorisée par l'autorité compétente du MESR.

Contraintes et risques

No specific risk identified