/Impact-of-Trajectory-Generation-Methods

Primary LanguagePython

Group Behavior Recognition Using Attention- and Graph-Based Neural Networks

[paper] [data]

If this code helps with your work, please cite:

@inproceedings{yang2020impact,
  title={Impact of trajectory generation methods on viewer perception of robot approaching group behaviors},
  author={Yang, Fangkai and Yin, Wenjie and Bj{\"o}rkman, M{\aa}rten and Peters, Christopher},
  booktitle={2020 29th IEEE International Conference on Robot and Human Interactive Communication (RO-MAN)},
  pages={509--516},
  year={2020},
  organization={IEEE}
}

Dataset

To explore our research questions, we designed an exploratory scenario, Who’s the Spy, for human-robot interactions.

This game involves three players in a group. In every game round, each player is given a card with a word on it. The Spy has a different word card. The players take turns to describe the word and the robot approaches to join the group to identify the spy.

network

Experimental Conditions

  1. Group type: a) static b) quasi-dynamic [1]
  2. Approaching Direction
  3. Camera viewpoints: a) egocentric view b) perspective view

network

Methods

  1. WoZ In the WoZ (Wizard-of-Oz) approach, the robot is teleoperated by a human operator (a researcher who is a trained operator).

network

  1. Procedural Model We use the social-aware navigation method as a procedural model to generate approaching group trajectories

network

  1. Imitation Learning Model We generate approaching group trajectories use a Generative Adversarial Imitation Learning (GAIL) framework with a Group Behavior Recognition framework.

network