Institut des Systèmes Intelligents
et de Robotique

Title : Doctorant.e
No longer in the unit

My work focuses on learning robot behaviors to interact with people. In order to do so, I've been involved in 2 types of studies:

• Group Modeling. The purpose is to understand how humans gather in public places and create an abstraction of a group of people; this abstraction is aimed to be used by a robot to understand certain behaviors of public gatherings in order to take decisions.
• Learning Robot Behaviors. The use of Inverse Reinforcement Learning (IRL) as a means to teach a robot how to approach a person and groups of people (current work).

This work is performed in partnership with the European project SPENCER. The scenario of SPENCER is deploying a robot in the Amsterdam airport being able to take into account human interactions.

Group Modeling

This study consists in creating a mathematical model to represent groups of people. Based on this information a robot could know whether or not it is included in a group of people, approach a group of people in a commercial center, or guide people to some gate at an airport. We plan to use the results in the future as heuristics for robots approaching groups of people.

I have developed two models, a static one and a dynamic one. The static is based on the Marco Cristani's work, nonetheless when trying to use it in a dynamic environment we failed to detect groups due to the lack of dynamic variables so we created a formalism in the dynamic case. The dynamic approach is an accepted paper to be published at IEEE International Symposium on Robot and Human Interactive Communication (RO-MAN).

Static Environment

Dynamic Environment

This model was created to tackle static and dynamic environments. It is based on low level features (positions and velocities in time) and uses novel techniques in order to segregate people in the environment (simulated and real). The first video is taken with  data gathered with an OptiTrack system, the second one with the SALSA database, and the third one with pedsim (ros).

Learning Robot Behaviors (To Approach People)

These studies aim to create navigation planners to find the best way to approach people. Using a combination of IRL with post-processing techniques, they guarantee optimal behaviors in terms of learning.

Scenario with one person

With a human centered map, which is then mapped in an MDP, we give demonstrations to a robot on how it's suppose to get close to a person. Then we apply IRL (based on Sergey Levine's Work) and its results are post-processed in order to create a path. This path can be considered as a type of global planner path for the navigation stack.

Another way to tackle this problem was using layered costmaps as seen in the following video.

Multiple People

This is work in progress using an MDP with generalized state-action  pairs and IRL.

Publications