VELMA: Verbalization Embodiment of LLM Agents for Vision and Language Navigation in Street View

TL;DR

This paper tackles urban vision-and-language navigation by grounding large language models in a real-world Street View environment. It introduces VELMA, a verbalization-based agent that reasons over a prompt containing the task, instructions, trajectory, and verbally encoded visual observations, including landmark visibility scored via CLIP and a landmark extractor. The approach yields strong results in both few-shot and finetuned settings, achieving state-of-the-art task completion on urban VLN benchmarks and showing that verbalized embodied guidance can enable LLMs to operate effectively in complex real-world navigation. The methods offer a scalable path to embodied reasoning with LLMs and highlight the importance of environment design, landmark grounding, and flexible action vocabularies like TURN_AROUND for robust navigation.

Abstract

Incremental decision making in real-world environments is one of the most challenging tasks in embodied artificial intelligence. One particularly demanding scenario is Vision and Language Navigation~(VLN) which requires visual and natural language understanding as well as spatial and temporal reasoning capabilities. The embodied agent needs to ground its understanding of navigation instructions in observations of a real-world environment like Street View. Despite the impressive results of LLMs in other research areas, it is an ongoing problem of how to best connect them with an interactive visual environment. In this work, we propose VELMA, an embodied LLM agent that uses a verbalization of the trajectory and of visual environment observations as contextual prompt for the next action. Visual information is verbalized by a pipeline that extracts landmarks from the human written navigation instructions and uses CLIP to determine their visibility in the current panorama view. We show that VELMA is able to successfully follow navigation instructions in Street View with only two in-context examples. We further finetune the LLM agent on a few thousand examples and achieve 25%-30% relative improvement in task completion over the previous state-of-the-art for two datasets.

Figures (10)

• Figure 1: Prompt sequence used to utilize LLMs for VLN in Street View. Verbalized observations of the visual environment are in green and appended to the prompt at each step. Agent actions (blue) are acquired by LLM next word prediction. Highlighting of text for visual presentation only. Full navigation trajectories are, on average, 40 steps long.
• Figure 2: The Touchdown environment introduced by Chen2018Touchdown can require action sequences that are semantically inconsistent with the correct navigation instructions. In the depicted subgraph, the action sequence to move from node 1 to node 5 is to move ${\tt FORWARD}$ four times. The semantically correct sequence of actions would include a right turn in between. We fix the problem by modifying the environment behavior and selecting the desired direction at intersections in relation to all outgoing streets.
• Figure 3: Overview of the proposed agent VELMA navigating in the Street View environment. The prompt sequence includes the task description, navigation instructions, and verbalized navigation trajectory up to the current timestep. The next action is decided by next word prediction utilizing an LLM and subsequently executed in the environment. This puts the agent into a new state, and the landmark scorer determines if an extracted landmark is visible in the current panorama view. The verbalizer takes this landmark information along with the information about a potential intersection and produces the current observations text. This text is then appended to the prompt sequence and again used to predict the next action. This process is repeated until the agent stops and the alleged target location.
• Figure 4: Distribution of CLIP scores between a landmark and panorama images in the training area. The CLIP score represents the semantic similarity of the panorama image and the text caption "picture of $[$landmark$]$". The distribution is used to standardize the score of the landmark and a novel panorama. The threshold $\tau$ is defined on the standardized score and used to determine the visibility of the landmark in the novel panorama image.
• Figure 5: Key point accuracy (KPA) for 2-shot in-context learning of large language models with increasing parameter count. The ${\tt FORWARD}$-Only baseline predicts walking forward until the average trajectory length is reached and performs better than predicting random directions.