EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Long-Horizon Humanoid Robot Vehicle Interaction via Video-Based Motion Priors and Imitation Learning

Hongming Chen, Zhen Xu, Yucheng Meng, Kenji Hashimoto

Biomimetics · 15.11.2026 · 11 (11) · 833

DOI: 10.3390/biomimetics11110833 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Biomimetics, online erschienen am 15.11.2026. Verfasst von Hongming Chen und 3 weiteren Autor:innen. Die Arbeit nennt 40 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: JSPS KAKENHI, Kitakyushu Foundation for the Advancement of Industry, Science and Technology, Waseda University.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Humanoid robots working in human-centered environments must perform navigation and contact-rich interaction continuously over long horizons. Vehicle interaction captures this requirement in a compact form: the robot has to approach a vehicle along a planned route, enter it and reach the seated target, and then leave the vehicle and continue toward a new destination. Such a capability serves practical needs ranging from disaster response and emergency rescue to daily assistance and can be regarded as an embodied basis for more complex human–robot collaboration and future vehicle-operation tasks. This paper presents a framework for long-horizon humanoid robot vehicle-interaction motion generation based on video motion priors and imitation learning. We first obtain multiple reusable vehicle-interaction and locomotion skills through human motion priors and task-conditioned adversarial imitation learning and adopt a task-conditioned Transformer policy so that a single policy can learn and execute different types of robot skills. A staged training strategy then learns the complete long-horizon task by jointly training complete task sequences, adjacent skill transitions, and individual skills, together with a policy adapter module, which improves the reliability of continuous multi-skill composition and the stability of training. Simulation experiments demonstrate reliable long-horizon vehicle-interaction execution across newly sampled configurations and externally planned routes, using simulator-provided geometry and task information in environments with static obstacles.

Förderung

  • JSPS KAKENHI
  • Kitakyushu Foundation for the Advancement of Industry, Science and Technology
  • Waseda University
  • JST SPRING

Zitieren

APA

Chen, H., Xu, Z., Meng, Y., & Hashimoto, K. (2026). Long-Horizon Humanoid Robot Vehicle Interaction via Video-Based Motion Priors and Imitation Learning. Biomimetics, 11(11), 833. https://doi.org/10.3390/biomimetics11110833

Vancouver

Chen H, Xu Z, Meng Y, Hashimoto K. Long-Horizon Humanoid Robot Vehicle Interaction via Video-Based Motion Priors and Imitation Learning. Biomimetics. 2026;11(11):833. doi:10.3390/biomimetics11110833

BibTeX
@article{Chen2026,
  author = {Chen, Hongming and Xu, Zhen and Meng, Yucheng and Hashimoto, Kenji},
  title = {Long-Horizon Humanoid Robot Vehicle Interaction via Video-Based Motion Priors and Imitation Learning},
  journal = {Biomimetics},
  year = {2026},
  volume = {11},
  number = {11},
  pages = {833},
  doi = {10.3390/biomimetics11110833}
}
Kostenlos starten →