Two papers accepted at MIG2020:

  1. Haworth B, Berseth G, Moon S, Faloutsos P, Kapadia M. Deep Integration of Physical Humanoid Control and Crowd Navigation.
  2. Kremer M, Haworth B, Kapadia M, Faloutsos P. Watch Out! Modelling Pedestrians with Egocentric Distractions.

Deep Integration of Physical Humanoid Control and Crowd Navigation brings together for the first synthetic crowds and physical character control. We show that it is possible to achieve extremely high fidelity crowds and learn anticipatory reciprocal collision avoidance using hierarchical multi-agent reinforcement learning. This work represents several years of work toward improving the fidelity and representation of synthetic crowds while improving the affordances of our past modelling approach. This approach promises to increase the representation and diversity of the humans we model.

Watch Out! Modelling Pedestrians with Egocentric Distractions models modern crowd agents by smartly combining parametric and event-driven behaviour trees on top of an agent-based distraction model. We introduce a parametric fuzzy goal notion, to model under/overshoot and lateral goal deviation, and a visual attention field of Gaussians combined in polar coordinates, to model the probability of noticing external objects during egocentric distractions. This work represents a significant increment to our previous work showing that distracted cell phone use can have a significant impact on essential crowd measures like flow rate (Modelling distracted agents in crowd simulations).