<p>In this study we present ETPNav as a robust vision-language navigation (VLN) system that uses a combination of an online topological map with a transformer model for predicting waypoints based on natural language instructions in order to create a reliable method for following instructions in complex and changing indoor environments. In contrast to many other VLN models that are trained using static representations of scenes, ETPNav learns a graph of how spaces are related by constructing it incrementally using RGB-D data during exploration. This enables ETPNav to reason about the relationships between known spaces and unknown spaces in its environment. The high-level planner is a transformer that predicts waypoints that correspond to where the agent should move in order to reach a goal, and the low-level controller generates smooth motion commands to guide the agent through the environment based on odometry and sensor readings. We evaluate ETPNav using two benchmarks that test both the ability of the model to follow instructions and the robustness of the model to changes in the environment. We find that ETPNav performs significantly better than prior methods that use either waypoints or transformers, and that it maintains its performance across all three of our test conditions: noisy odometry, ambiguous language, and moving objects in the environment. Finally, we provide qualitative examples that demonstrate that ETPNav produces smoother paths and is more robust than previous models. Although we have made several improvements to prior methods, ETPNav still relies on simulated RGB-D data and assumes that the agent’s motion is planar. Therefore, future work will be necessary in order to adapt ETPNav to an embodied agent that can perceive its surroundings and create maps of real-world environments.</p>