Original title: Revisiting Topological Graphs for Macro Action based Closed-loop Reinforcement Learning of Vision Language Navigation in Continuous Environment
Shuhao Ye, Sitong Mao, Yuxiang Cui, Yufei Wei +6 · September 2026
Robots struggle to follow spoken directions across unfamiliar rooms. Teaching them by copying human moves often fails when the robot wanders off track. Trying random trial-and-error actions usually takes too long because the robot gets lost making tiny movements. Researchers solved this by splitting navigation into two parts: high-level planning and low-level steering. They turned spaces into a map of connected points called a topological graph. Instead of choosing every tiny step, the robot picks a major destination point on this map. A controller guides it there without extra training. They also built an action-aware value tool to score these decisions during training. This approach makes training faster and helped the robot reach state-of-the-art navigation performance across two benchmark tests.
Imagine you're giving a friend directions to your house over the phone. If you say 'turn left after the red building,' but your friend doesn't know what a red building is, they might get lost immediately. In this paper, the system uses a simplified map with key landmarks to plan routes, so even if the environment is unfamiliar, it can navigate by following high-level steps to those landmarks.
This matters because robots that can understand and follow spoken directions in homes or offices could assist elderly or disabled people more reliably. Right now, they often fail when environments are new or instructions are complex.
The abstract does not state any limitations; it only reports success on benchmarks R2R-CE and RxR-CE.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Revisiting Topological Graphs for Macro Action based Closed-loop Reinforcement Learning of Vision Language Navigation in Continuous Environment — Shuhao Ye, Sitong Mao, Yuxiang Cui 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.