SG-LSTM: Social Group LSTM for Robot Navigation Through Dense Crowds

Rashmi Bhaskara,Maurice Chiu,Aniket Bera

2023-08-07

Abstract:With the increasing availability and affordability of personal robots, they will no longer be confined to large corporate warehouses or factories but will instead be expected to operate in less controlled environments alongside larger groups of people. In addition to ensuring safety and efficiency, it is crucial to minimize any negative psychological impact robots may have on humans and follow unwritten social norms in these situations. Our research aims to develop a model that can predict the movements of pedestrians and perceptually-social groups in crowded environments. We introduce a new Social Group Long Short-term Memory (SG-LSTM) model that models human groups and interactions in dense environments using a socially-aware LSTM to produce more accurate trajectory predictions. Our approach enables navigation algorithms to calculate collision-free paths faster and more accurately in crowded environments. Additionally, we also release a large video dataset with labeled pedestrian groups for the broader social navigation community. We show comparisons with different metrics on different datasets (ETH, Hotel, MOT15) and different prediction approaches (LIN, LSTM, O-LSTM, S-LSTM) as well as runtime performance.

Robotics,Artificial Intelligence,Multiagent Systems

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to achieve social navigation of robots in dense crowds. As personal robots become more and more popular and affordable, their application scope has expanded from large - scale enterprise warehouses and factories to more diverse and less - controlled environments, such as public places. In these environments, robots need not only to ensure safety and efficiency, but also to mitigate potential adverse psychological impacts on humans and abide by unwritten social norms. To meet these challenges, the paper proposes a new method named "Social Group Long Short - term Memory" (SG - LSTM), aiming to predict pedestrian movement and interaction, especially in crowded environments. By integrating social awareness into the LSTM architecture, the SG - LSTM model can predict trajectories more accurately, thus helping the navigation algorithm calculate collision - free paths more quickly and accurately, especially in complex and crowded situations. In addition, the author also provides a large - scale video data set containing marked pedestrian groups to promote the further development of social navigation research.