End-to-end Learning of Driving Models from Large-scale Video Datasets
This addresses the need for more generalizable driving models that can handle diverse visual conditions, moving beyond limited in-situ or simulated training data.
The paper tackles the problem of learning robust perception-action models for autonomous driving by developing an end-to-end trainable architecture that predicts future vehicle egomotion from monocular camera observations and previous state, using a novel FCN-LSTM model trained on large-scale crowd-sourced video data.
Robust perception-action models should be learned from training data with diverse visual appearances and realistic behaviors, yet current approaches to deep visuomotor policy learning have been generally limited to in-situ models learned from a single vehicle or a simulation environment. We advocate learning a generic vehicle motion model from large scale crowd-sourced video data, and develop an end-to-end trainable architecture for learning to predict a distribution over future vehicle egomotion from instantaneous monocular camera observations and previous vehicle state. Our model incorporates a novel FCN-LSTM architecture, which can be learned from large-scale crowd-sourced vehicle action data, and leverages available scene segmentation side tasks to improve performance under a privileged learning paradigm.