Prediction multimodale pour robots en interaction avec des humains

Mots clés

machine learning, robotique

Détail de l'offre

The objective of this Ph.D. is to propose new algorithms to accurately predict the future whole-body motion of a human, which is critical for human-robot collaboration. This has a direct application in social robots co-existing or interacting with humans in their environment, in exoskeletons, and in the direct teleoperation of robots with different degrees of anthropomorphism.
The thesis will target several advances in human motion prediction, such as predicting the intention of motion at different time scales, with different families of models, adding physical priors to guarantee the prediction follows the model of the human and physics constraints (e.g., the human is balanced, with the feet on the floor), and incorporating external sources of information to improve the prediction while taking the environment into account (objects nearby, payloads, obstacles from RGB-D sensors, etc.).
The level of physical interaction in this thesis will be minimal, as the focus will be particularly on situations when the onset of action must be detected to ensure a prompt anticipatory action from the robot. As an example, handovers or collaborative lifting/carrying will be studied.
The software and methods produced by Ph.D. will be a basic “building block' for many interaction scenarios addressed in the PEPR project: developing open software that can be re-used in other projects will be of paramount importance.