UESF-Bench: Benchmarking and Probing for Unified Embodied Seeking and Following

This paper introduces UESF-Bench, a benchmark for embodied agents that must first locate a language-described person and then follow that person persistently in dynamic environments. It evaluates semantic-guided exploration, switching between seeking and following, recovery after failures, and delayed identity grounding. The authors also propose SeekFollow-VLA, a vision-language-action framework with task-driven routing for latent phase inference and transition modeling. According to the abstract, it outperforms single-head and dual-head baselines across single-person and multi-person settings, providing a unified baseline for seek-and-follow behavior.