Model-based Imitation Learning for Real-time Robot Navigation in Crowds
Martin Moder, Fatih Özgan, Josef Pauli · 2023
We are increasingly interacting with robots in our everyday life. To further this development, a critical capability is safe and socially compliant robot navigation in a crowd. In this work, we extract a navigation strategy from past human-to-human interactions with a model-based approach to imitation learning. We propose a hybrid model of crowd dynamics that combines an autoregressive and an inverse autoregressive model for real-time sampling-based planning with respect to human decisions making. Furthermore, we constrain the optimization to allow only admissible velocities for any given robot dynamics that lead to a trajectory on which the robot can safely stop. Experiments are conducted in crowded prerecorded environments, where the robot is placed in a variety of scenarios with varying numbers of humans. The results show that the algorithm is able to navigate in these environments with a lower collision rate and a shorter path than the state-of-the-art.