New Benchmark for Embodied AI Human Seeking and Following

Kun Yu, Jianhua Yang, Yixiang Chen, Changwei Wang, Hongyuan Yu, Yan Huang, Fushuo Huo, Ya Jing, Zhumin Chen, Keji He· July 16, 2026 View original

Key takeaways

  • UESF-Bench offers a unified benchmark for embodied AI to seek and follow humans.
  • The benchmark addresses realistic scenarios requiring initial search and persistent following.
  • SeekFollow-VLA framework improves performance in dynamic, multi-person environments.
  • This research advances human-robot interaction and autonomous system capabilities.

Who benefits

RoboticsLogisticsHealthcareSmart HomesDefense

Summary

Researchers introduce UESF-Bench, a new large-scale benchmark for embodied agents to first find a language-described human target and then persistently follow them in dynamic environments, addressing limitations of prior task-specific scenarios.

A new research paper presents UESF-Bench, a novel benchmark designed to evaluate embodied AI agents in more realistic human interaction scenarios. Unlike previous benchmarks that often assume the target is immediately visible, UESF-Bench challenges agents to actively search for a language-described person before initiating a following behavior. This unified approach integrates semantic-guided exploration, reliable behavior switching, and delayed identity grounding, reflecting the complexities of real-world dynamic environments. To tackle this, the paper also proposes SeekFollow-VLA, a vision-language-action framework that uses a task-driven routing mechanism to manage transitions between seeking and following phases. Experimental results demonstrate that SeekFollow-VLA significantly outperforms existing baselines in both single and multi-person settings, establishing a new foundation for future research in this area.

Why it matters

This research is crucial for advancing the capabilities of embodied AI, enabling robots and autonomous systems to interact more naturally and effectively with humans in complex, real-world settings.

How to implement this in your domain

  1. 1Review the UESF-Bench framework for potential application in developing more robust human-robot interaction systems.
  2. 2Explore the SeekFollow-VLA architecture to inform the design of vision-language-action models for autonomous agents.
  3. 3Consider integrating semantic-guided exploration and dynamic behavior switching into existing robotic navigation or assistance platforms.
  4. 4Participate in the benchmark to test and validate new algorithms for embodied seeking and following tasks.

Original post by Kun Yu, Jianhua Yang, Yixiang Chen, Changwei Wang, Hongyuan Yu, Yan Huang, Fushuo Huo, Ya Jing, Zhumin Chen, Keji He

"arXiv:2607.13621v1 Announce Type: new Abstract: Language-guided human following is an important capability for embodied agents, but existing benchmarks typically assume that the target person is visible at the start of an episode. This setting simplifies the problem and overlooks…"

View on X

Originally posted by Kun Yu, Jianhua Yang, Yixiang Chen, Changwei Wang, Hongyuan Yu, Yan Huang, Fushuo Huo, Ya Jing, Zhumin Chen, Keji He on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses