José Grimaldo da Silva Filho - Towards natural human-robot collaboration during collision avoidance

14:00
Thursday
6
Feb
2020
Organized by: 
James L. Crowley
Speaker: 
José Grimaldo da Silva Filho

 

Jury :

  • Alberto Sanfeliu, professeur, Univ. Polytechnique de Catalogne Espagne, rapporteur
  • Rachid  Alami, directeur de recherche HDR, CNRS-LAAS, rapporteur
  • Olivier  Simonin, professeur, INSA Lyon, examinateur
  • Patrick  Reignier, professeur, Grenoble INP, examinateur
  • James L. Crowley, professeur, directeur de thèse

 

Classical approaches for robot navigation among people have focused on guaranteed collision-free motion with the assumption that people are either static or moving obstacles. However, people are not ordinary obstacles. People react to the presence and the motion of a robot. In this context, a robot that behaves in human-like manner has been shown to reduce overall cognitive effort for nearby people as they do not have to actively think about a robot's intentions while moving on its proximity.

Our work is focused on replicating a characteristic of human-human interaction during collision avoidance that is the mutual sharing of effort to avoid a collision. Based on hundreds of situations where two people have crossing trajectories, we determined how total effort is shared between agents depending on several factors of the interaction such as crossing angle and time to collision. As a proof of concept our generated model is integrated into rvo. For validation, the trajectories generated by our approach are compared to the standard rvo and to our dataset of people with crossing trajectories.

Collaboration during collision avoidance is not without its potential negative consequences. For effective collaboration both agents have to pass each other on the same side. However, whenever the decision of which side collision should be avoided from is not consistent for people, the robot should also account for the risk that both agents will attempt to incorrectly cross each other on different sides. Our work first determines the uncertainty around this decision for people. Based on this, a collision avoidance approach is proposed so that, even if agents initially choose to incorrectly attempt to cross each other on different sides, the robot and the person would be able to perceive the side from which collision should be avoided in their following collision avoidance action. To validate our approach, several distinct scenarios where the crossing side decision is ambiguous are presented alongside collision avoidance trajectories generated by our approach in such scenarios.