Recoleta
GitHub
Source note

Walk With Me: Long-Horizon Social Navigation for Human-Centric Outdoor Assistance

Authors
Lingfeng Zhang, Xiaoshuai Hao, Xizhou Bu, Yingbo Tang, Hongsheng Li, Jinghui Lu, et al.
Source
arXiv
Published
Collection
Embodied AI
Social NavigationOutdoor NavigationVision Language ActionHuman Robot AssistanceMap Free Navigation
Open arXivMarkdown source

Summary

Walk with Me is a robot navigation system for outdoor assistance from high-level language instructions. It grounds user intent with public map POIs, plans coarse waypoints, and uses VLM/VLA modules for local motion and safety decisions.

Problem

  • It solves long-horizon outdoor social navigation where a user gives an abstract request such as “I want to go for a walk” instead of a coordinate goal.
  • This matters for last-mile delivery and blind guidance, where the robot must choose a real destination, follow a long route, and behave safely near crossings, traffic, and crowds.
  • HD-map navigation is costly to build and maintain, while many learned navigation policies focus on indoor or short routes with low-level goals.

Approach

  • A High-Level VLM receives the user instruction, GPS context, and candidate POIs from a public map API, then selects a concrete destination.
  • The system queries a walking-route API and resamples the route into geo-referenced waypoints for long-horizon guidance.
  • At each step, the robot forms an observation from the RGB image, local pose, recent trajectory, and next waypoint.
  • A VLM router decides whether the scene is routine or complex and whether the robot should proceed or stop and wait.
  • If proceeding is safe, a Low-Level VLA predicts a short local trajectory; if safety confidence is low, the robot waits and rechecks the scene.

Results

  • The real-world evaluation uses 20 outdoor trials: 2 application categories, 2 scenarios per category, and 5 independent trials per scenario.
  • The two application categories are last-mile delivery and blind guidance; example instructions include “Take the milk tea to Building B” and “I want to go shopping.”
  • The paper claims kilometer-scale outdoor navigation from abstract instructions on an Athena 2.0 Pro AGV robot.
  • The excerpt gives no success rate, path length distribution, completion time, collision rate, or social-compliance metric.
  • The main experiment reports no full-system baseline comparison; the authors state that existing methods do not match the same input-output setting.

Link