Intervention Force-based Imitation Learning for Autonomous Navigation in Dynamic Environments

Imitation learning is a data-driven approach that has proven to be successful in building autonomous navigation systems. One of the key tasks in imitation learning is collecting data, but data only collected by humans cannot include many types of data, such as deviating from the target path. If we use a model trained with such data, the deviation will accumulate, and returning to the target path will be difficult. Related studies have demonstrated the importance of correction data, and two types of solutions have been proposed: data augmentation and online sampling. In this paper, we propose a new online sampling method for acquiring correction data that is safe and effective, which uses a device that detects the force applied a steering wheel and accelerator pedal during an intervention. Autonomous navigation experiments are conducted using small vehicles to follow a specified path in static and dynamic environments. Our experimental results show that we can successfully separate intervention data from all collected data using intervention force and that using intervention data for model training is effective for improving the route-following. Also, our model can perform obstacle avoidance and generate appropriate control signals when encountering dynamic objects, such as pedestrians, at specific locations.