RAIBO-Wagon: Real-Time Navigation System
for a Legged Robot Towing a Wagon

Insung Yang, Eunchae Kim, Jungwoo Hur, Seunghun Jeon, Donghyuk Choi, and Jemin Hwangbo*

Robotics and Artificial Intelligence Lab, KAIST

*Corresponding author

TL;DR

Towing a wagon lets a quadruped haul far more than it can carry, but the coupled system is underactuated, has an enlarged footprint, and the wagon can only roll along its own heading, so navigating confined spaces in real time is hard. RAIBO-Wagon tackles this by warm-starting MPPI with a Conditional Flow Matching (CFM) model distilled from an offline expert, then refining the samples with a learned Forward Dynamics Model (FDM), entirely onboard. In 40 real-world trials on 15 m courses with static obstacles and pedestrians, it succeeds 85% of the time while towing up to 40 kg, nearly RAIBO's own body weight.

Method

This video explains our method by walking through a single planning cycle step by step.

Results

Real-World Navigation

RAIBO-Wagon drives a 15 m outdoor course with no prior map, through six static obstacle layouts (Terrains A–F) and two where pedestrians block the route (G, H), five trials each. It reaches the goal in 34 of 40 trials.

Terrain

Additional Mass Robustness

We load the ~25 kg wagon with an extra 5, 10, or 15 kg on Terrain F, five trials each, without telling the system the mass. It succeeds in every trial.

Baseline Study

We compare our framework with Hybrid A* and an RL navigation policy on Terrain F, five trials each. Both baselines fail at the narrowest corridor, while ours reaches the goal every time.

BibTeX

@article{yang2026raibowagon,
  title   = {{RAIBO-Wagon}: Real-Time Navigation System for a Legged Robot Towing a Wagon},
  author  = {Yang, Insung and Kim, Eunchae and Hur, Jungwoo and Jeon, Seunghun and Choi, Donghyuk and Hwangbo, Jemin},
  journal = {arXiv preprint},
  year    = {2026}
}