Original title: Air-Ground Collaborative Vision-and-Language Navigation via Shared Bird's-Eye Maps
Shuning Zhang, Liang Li, Yunheng Wang, Tao Wang +2 · September 2026
Navigating through an area using vision and language instructions is hard for a single robot. Flying drones see the whole picture from above, while ground rovers only see what is directly ahead. Previous attempts to make them work together either failed to create stable cooperation or actually made performance worse. To fix this, researchers created AGC-VLNVision-and-Language Navigation, a task where a robot follows natural language instructions using visual input., a system where the flying drone builds a shared overhead map. It marks the ground rover's location, the target goal, and distances directly on this view. The ground rover uses this shared map to see the bigger picture, plan a route along roads, and drive along it. At the same time, the drone searches downward in three dimensions to find the goal and fly toward it. In tests across 100 trials, the team reached a 77.0% joint success rate, showing that overhead vision and ground driving complement each other well.
Imagine two people setting up a tent in a forest. One stands on a hill with a map of the campsite, while the other is in the clearing and can only see nearby trees. The person on the hill knows where the tent should go, but without a shared view, directions get mixed up. In this paper, the drone acts like the person on the hill, sharing a bird's-eye map with the ground robot, so it knows the exact layout and can navigate without confusion.
This helps robots coordinate better without needing to learn from scratch. For tasks like disaster response, a drone and ground vehicle could work together more reliably, reducing human effort.
The method is only tested in a single simulated environment (CARLA-Air's Town10HD) with 100 episodes, so it might not work in different settings. It relies on specific components like frozen VLMs and 3D-SPF, which could limit adaptability. As a training-free baseline, it doesn't learn from experience and might be outperformed by trained systems.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Air-Ground Collaborative Vision-and-Language Navigation via Shared Bird's-Eye Maps — Shuning Zhang, Liang Li, Yunheng Wang et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.