Vous êtes ici

Crowd simulation for safe robot navigation in public places

Equipe et encadrants
Département / Equipe: 
Site Web Equipe: 
Directeur de thèse
Julien Pettre
Co-directeur(s), co-encadrant(s)
NomAdresse e-mailTéléphone
Julien Pettre
Sujet de thèse

General Context

The context of this PhD thesis is the H2020 CROWDBOT project, the objective of which is to make robots capable of navigating crowded environment. Moving a robot in a crowd means moving at close distance to humans in a highly dynamic environment, the risk of collision with humans is constantly high. Today, are programmed to stop when getting close to humans, preventing them to achieve any navigation task in a crowded environment. In addition, the stopped robots turn into an obstacle to the crowd motion, also provoking risks for people around the robot. For this reason, we aim at exploring alternative strategies to make robots capable of safely moving in a crowd. More specifically we want to provide the robot with new capabilities:

  • To perceive the crowd around the robot

  • To predict the evolution of the crowd around the robot

  • To perform safe navigation in this dynamic environment based on this prediction.

The specific objective of this PhD is to couple the perception component of the robot, i.e., the tracking of humans and obstacles around the robot, with the prediction component of the robot, i.e., a crowd simulator which predicts how humans move in a crowded context.


Crowd simulation techniques divide into two categories: the macroscopic approaches model the crowd as a whole, a continuous moving matter [Treuillle 2006], whereas microscopic approaches model each individual of a crowd as a multi-agents system. In this topic, we are interested in the latter category of approach. Recent progress in microscopic crowd modeling allowed simulations to reach new levels of realism [Wolinski 2016][Dutra 2017]. Our idea is to benefit from these recent progress in crowd simulation to provide robots with the ability to predict the evolution of a crowd of people around them. To this end, our key idea is to initialize a crowd simulation based on the robot perception of the crowd around it, and to run the simulation to dispose of a short term prediction of the crowd state evolution. This will serve as a basis for the robot to control its own navigation.

Several problems however need addressing:

  1. Microscopic crowd simulation algorithms are based on the notion of local model of interactions. Such models describe how one’s motion is influenced by the presence of neighbors, e.g., to achieve collision avoidance. Existing models of interactions do not consider interactions between robots and humans yet, and need to be revisited to correctly simulate how the neighbors of the robot will react to it. [Vassallo 2017] will be a starting point to address this issue.

  2. Simulation parameters greatly influence simulation results and need to be correctly estimated from the robot perception, in order for the simulation to fit the crowd actually observed by the robot. This concerns individual states (positions, velocities) but also less accessible parameters, such as prefered walking speed, goals and intentions, social relation between crowd members, etc. Our previous work on this topic will serve as a basis [Wolinski 2014][Bera 2017].

  3. The simulation update scheme need to be defined : update frequencies and simulation reset from robot perception, re-estimation and fusion of simulation parameters from the robot observation, etc.

  4. Evaluation of simulation accuracy and incertainty. It is crucial to provide guarantee about the quality of the simulation while the robot use the simulation results as a basis for safe navigation. A continuous estimation of the simulation accuracy will be possible from robot observation of the evolution of the crowd state.

Research environment

Position The position is opened at Inria in Rennes, France.

The net salary is 1.600 euros / month.

This position is opened in the context of the H2020 CROWDBOT European project. The candidate will have the opportunity of close collaboration with the project partners from EPFL, ETHZ, RWTH, UCL, SoftBank Robotics and Locomotec.


We are seeking extremely motivated candidates, with strong technical and scientific background. The following skills are appreciated:

- excellent developper (C++)

- background in machine learning is desired

- knowledge in computer vision, crowd simulation, robot navigation will be considered

- interest in human behavior simulation

Please send directly your application to julien.pettre@inria.fr



[Treuillle 2006] Treuille, A., Cooper, S., & Popović, Z. (2006, July). Continuum crowds. In ACM Transactions on Graphics (TOG) (Vol. 25, No. 3, pp. 1160-1168). ACM.

[Wolinski 2014] Wolinski, D., J Guy, S., Olivier, A. H., Lin, M., Manocha, D., & Pettré, J. (2014, May). Parameter estimation and comparative evaluation of crowd simulations. In Computer Graphics Forum (Vol. 33, No. 2, pp. 303-312).

[Wolinski 2016] Wolinski, D., Lin, M. C., & Pettré, J. (2016). WarpDriver: context-aware probabilistic motion prediction for crowd simulation. ACM Transactions on Graphics (TOG), 35(6), 164.

[Vassallo 2017] Vassallo, C., Olivier, A. H., Souères, P., Crétual, A., Stasse, O., & Pettré, J. (2017). How do walkers avoid a mobile robot crossing their way?. Gait & posture, 51, 97-103.

[Dutra 2017] Dutra, T. B., Marques, R., Cavalcante‐Neto, J. B., Vidal, C. A., & Pettré, J. (2017, May). Gradient‐based steering for vision‐based crowd simulation algorithms. In Computer Graphics Forum (Vol. 36, No. 2, pp. 337-348).

[Bera 2017] Bera, A., Wolinski, D., Pettré, J., & Manocha, D. (2014). Real-time crowd tracking using parameter optimized mixture of motion models. arXiv preprint arXiv:1409.4481.

Début des travaux: 
Mots clés: 
crowd simulation, robot navigation
IRISA - Campus universitaire de Beaulieu, Rennes