ReferTrack proposes a “referring-then-tracking” paradigm for embodied visual tracking with a single forward-facing camera. The model first selects a language-referred target from an indexed set of image bounding boxes, then predicts tracking waypoints conditioned on that grounded choice. A sliding queue of historical boxes is encoded through temporal-viewpoint-bbox indicator (TVBI) tokens to preserve target motion cues. The paper reports EVT-Bench success rates of 89.4%, 73.3%, and 74.1% on single-target, distracted, and ambiguity splits, respectively, and describes real-world deployments on legged and humanoid robots.
No heat snapshots are available in the last 24 hours.