Maps of Dynamics

To achieve the smooth, safe, and reliable operation of robots in proximity to humans, it is essential to enable robots to behave in a predictable, safe, and legible manner while adhering to common social norms. This requires robots to understand and follow general patterns governing the flow of people, ensuring socially acceptable behaviour, and anticipating human actions, thereby ensuring safe and efficient operation.

Over the past three decades, there has been significant progress in the development of human motion prediction algorithms, which allow robots to anticipate human actions in the near future. This capability enables robots to proactively avoid collisions, facilitating smooth and safe interactions between humans and robots. Additionally, recent years have seen rapid advancements in Maps of Dynamics (MoD), which capture global motion patterns governing human behaviour in specific environments. Although these two areas address different aspects of the same problem, they are still awaiting proper integration to fully utilize MoD in conjunction with motion prediction algorithms.

In this project, we aim to develop a novel framework that utilizes robotics mapping and representation learning techniques to address the modelling of dynamic aspects and human motion prediction problems. By collecting and modelling real-world pedestrian data, a long-term human motion prediction in environments with limited data access will be accomplished.

Keywords: 

Machine Learning; Deep Learning; Spatial Models of Dynamics; Human-Motion Prediction. 

People involved:

• Tomasz Piotr Kucner, [email protected]
• Junyi Shi, [email protected]