Visual Prompt Based Reasoning for Offroad Mapping using Multimodal LLMs

TL;DR

This paper introduces a zero-shot off-road navigation method using multimodal LLMs, combining environment segmentation and reasoning to identify drivable areas without task-specific training.

Contribution

It presents a unified, zero-shot framework leveraging SAM2 and vision-language models for off-road reasoning, eliminating the need for separate terrain models.

Findings

Outperforms state-of-the-art trainable models on segmentation datasets.

Enables full stack navigation in simulation environment.

Reduces reliance on task-specific datasets and fine-tuning.

Abstract

Traditional approaches to off-road autonomy rely on separate models for terrain classification, height estimation, and quantifying slip or slope conditions. Utilizing several models requires training each component separately, having task specific datasets, and fine-tuning. In this work, we present a zero-shot approach leveraging SAM2 for environment segmentation and a vision-language model (VLM) to reason about drivable areas. Our approach involves passing to the VLM both the original image and the segmented image annotated with numeric labels for each mask. The VLM is then prompted to identify which regions, represented by these numeric labels, are drivable. Combined with planning and control modules, this unified framework eliminates the need for explicit terrain-specific models and relies instead on the inherent reasoning capabilities of the VLM. Our approach surpasses state-of-the-art trainable models on high resolution segmentation datasets and enables full stack navigation in our Isaac Sim offroad environment.