Embodied Drone Tracking in the Wild: Benchmark and Baseline

Embodied drone tracking requires an agent to follow a moving target while actively controlling its viewpoint and avoiding obstacles. Drones offer broad views and flexible altitude adjustment, yet target motion and interrupted visibility make this coupling challenging. Existing benchmarks provide limited coverage of substantial target elevation changes and complex occlusions under a common promptable tracking protocol. We introduce the Embodied Drone Tracking Benchmark (EDT-BENCH), built on UnrealZoo to address this gap. It supports language, point, bounding box, mask, and template initialization across diverse environments. The training corpus contains 5,670 expert trajectories and their perturbed counterparts from 81 scenes, providing aligned supervision for tracking and recovery. We further propose EDT, a simple baseline built on frozen SAM2 representations. Its action head estimates relative target pose from current observations and temporal memory, then uses uncertainty-aware fusion to predict drone waypoints. The memory branch retains target information during occlusion, while spatial prompts reduce ambiguity when similar distractors are present. A unified closed-loop protocol evaluates tracking robustness and generalization across scenes and target categories. Together, EDT-BENCH and EDT provide a reproducible closed-loop framework for evaluating prompt-conditioned tracking, spatial tracking accuracy, recovery from occlusion and viewpoint perturbations, and generalization across scenes and target categories.