KT-Imitation: Efficient Visual Imitation for Autonomous Navigation Based on Keypoint Timeline

We introduce Keypoint Timeline (KT-) Imitation, a vision-only imitation algorithm for autonomous navigation. The algorithm is both data- and computation-efficient, as it only requires a single video demonstration. Different from methods in the literature, it does not require extensive action-observation information or an iterative learning process. The algorithm creates a timeline from a video demonstration to evaluate and fine-tune parameters for both the demonstrations and motion controls. This optimization improves the reliability and continuity of the ORB keypoints used in imitation tasks. We present an egocentric action-generation procedure that guides the imitator to carry the camera autonomously to follow the demonstration using the detected keypoints of the onboard camera. Our experiments validate the proposed approaches on a ground robot and an aerial vehicle against ground-truth recordings.