AIAny
Icon for item

ReferTrack: Referring Then Tracking for Embodied Visual Tracking

Selects a referred target from candidate bounding boxes, then decodes tracking waypoints for single-camera embodied visual tracking. Injects past selected-bbox geometry via sliding-window TVBI tokens and is co-trained on a Refer‑QA dataset; achieves SOTA on EVT‑Bench and demonstrates sim-to-real on legged and humanoid robots.

Introduction

Embodied visual tracking demands both reliable target identification from a natural-language cue and closed-loop trajectory reasoning under onboard vision. Many vision-language-action (VLA) approaches fold identification into abstract spatial latents, making supervision and image-space validation difficult. ReferTrack takes the opposite tack: make the referring decision explicit in image space first, then condition waypoint generation on that grounded choice. This simple decomposition yields clearer supervision, preserves appearance/motion cues, and improves identification-heavy tracking performance.

Key Findings
  • Explicit refer-then-track decomposition: the model first picks the referred target from an indexed set of bounding boxes, then decodes tracking waypoints conditioned on that image-grounded selection — this reduces failure modes caused by latent ambiguity in end-to-end VLA pipelines.
  • Temporal-viewpoint-bbox indicator (TVBI) tokens: a sliding-window queue of previously selected bounding boxes is encoded and injected into the visual history, retaining geometric and motion cues across timesteps so the tracker keeps sight of target motion patterns.
  • Data and co-training: co-training on a custom Refer‑QA dataset sharpens identification under ambiguity and distraction, improving robustness when multiple similar objects are present.
  • Empirical results: on EVT‑Bench ReferTrack achieves 89.4% / 73.3% / 74.1% success on the single-target, distracted, and ambiguity splits, matching or outperforming several multi-camera baselines on identification-heavy tasks. The method also transfers to real legged and humanoid robots in sim-to-real tests.
Who it's for and tradeoffs

Great fit if you build embodied agents that must follow language-specified targets using a single forward camera and you care about reliable identification under clutter or ambiguity. The architecture is especially relevant when image-space interpretability and debugging of the identification step matter. Look elsewhere if your system already has multi-view sensing, explicit external trackers, or constraints that preclude running a bbox-indexing frontend; ReferTrack focuses on image-grounded referring plus waypoint decoding rather than multi-sensor fusion or low-level locomotion control.

How it works (brief)

ReferTrack indexes candidate detections into a short-term memory and runs a referring module to select the target bbox. The selected bbox stream over time is summarized via TVBI tokens injected into the visual history encoder. A waypoint decoder then generates future motion waypoints conditioned on the grounded selection and visual history. Co-training with Refer‑QA improves the referring module's accuracy under semantic ambiguity. The authors release code and report both simulation benchmarks and robot deployments demonstrating sim-to-real transfer.

Information

  • Websitearxiv.org
  • AuthorsHanjing Ye, Tianle Zeng, Jiazhao Zhang, Shaoan Wang, Zibo Zhang, Weisi Situ, Yuchen Zhou, Yonggen Ling, Hong Zhang
  • Published date2026/07/22

More Items

Evaluates spatial cognition of image-generation models by eliciting protocol-constrained visual answers and parsing pixel outputs into structured predictions compatible with existing metrics. Introduces the ProVisE framework and SpatialGen-Bench (470 samples) to compare image-generation models and text-output VLMs on unified spatial tasks.

Converts image-content removal into a contrastive on-policy self-distillation signal: the EMA teacher produces next-token distributions with and without image content, uses their log-probability differences to sharpen visual-grounded candidates, and distills that full-distribution target into the student—no external teacher or extra inference cost.

Quantifies active visual observation in multimodal LLMs with ActiveVision, a 17-task benchmark that forces repeated perception rather than one-shot description. Finds frontier MLLMs fail badly (top model 10.6% vs humans 96.1%) and that model-generated vision code does not close the gap.