TLDR: DIV-Nav is a novel real-time navigation system that enables robots to understand and execute complex natural language commands involving spatial relationships between multiple objects. It achieves this by decomposing instructions using a Large Visual-Language Model (LVLM), combining semantic belief maps to identify regions where objects are likely to co-exist, and then validating the spatial constraints with the LVLM. The system demonstrated superior performance in both simulated and real-world multi-object navigation tasks, significantly improving success rates by effectively incorporating spatial reasoning.
Robots are becoming increasingly capable, but understanding complex human commands remains a significant challenge. While a robot might easily find a “chair,” asking it to “find the remote on the table” introduces a layer of spatial reasoning that traditional navigation systems often struggle with.
A new research paper introduces DIV-Nav, a real-time navigation system designed to bridge this gap. DIV-Nav empowers robots to interpret and act upon free-text queries that include intricate spatial relationships between multiple objects, all while leveraging robust semantic maps of their environment.
The Challenge of Spatial Relationships
Current zero-shot object navigation methods typically focus on simple object identification. They often create representations specific to a single query, failing to capture relationships like “next to” or “inside.” Even advanced Vision-Language Models (VLMs) and Large Visual-Language Models (LVLMs) like GPT5, while impressive in understanding images, don’t inherently possess the robust spatial and motion understanding needed to directly steer a robot based on complex spatial instructions.
This is where DIV-Nav steps in. It combines the reasoning power of LVLMs with incremental semantic exploration and mapping to enable robots to perform sophisticated searches.
How DIV-Nav Works: Decompose, Intersect, Validate
The core of DIV-Nav lies in a three-step process:
1. Decomposition: When given a complex natural language instruction (e.g., “find the remote on the table”), an LVLM module breaks it down into simpler, object-level queries. It identifies all mentioned objects and locations, infers higher-level concepts (like “bathroom” for “towel”), and determines the primary target along with its spatial relationship to other objects. For instance, “the blue rug in the bathroom” would be decomposed into “blue rug” as the primary target and “bathroom” with an “in” (proximity) relationship.
2. Intersection: The system then queries a semantic belief map with these individual object targets. This map, built incrementally as the robot explores, provides similarity values for each object across the environment. DIV-Nav combines these individual similarity maps to create a “joint similarity map.” This map highlights regions where all specified objects are likely to co-exist, effectively identifying areas that satisfy the proximity-based spatial relationships.
3. Validation: Once the robot navigates to a high-similarity region, it uses the LVLM again for visual inference. This step validates whether the discovered objects truly match the original, complex spatial constraints. For example, it confirms not just the presence of a “remote” and a “table,” but also that the “remote” is indeed “on” the “table.” If the validation fails, the search continues.
Also Read:
- FALCON: Improving Robot Dexterity Through Advanced Spatial Perception
- Optimizing Vision-Language-Action Models for Robotics: A Deep Dive into Efficiency
Guiding Exploration and Real-World Performance
DIV-Nav adapts frontier exploration objectives to guide the robot more effectively. It scores potential exploration frontiers and previously observed areas based on the combined similarity map, directing the robot towards locations where the target objects and their spatial relationships are most likely to be found.
The system was rigorously tested on the MultiON benchmark and deployed on a Boston Dynamics Spot robot. In simulated environments, DIV-Nav showed a 30% improvement in success rate per attempted target (SRAT) and a 40% improvement in overall progress compared to baselines that lacked spatial reasoning capabilities. In real-world experiments across various indoor scenes (office, lab, kitchen), DIV-Nav achieved an impressive 88% success rate, successfully finding 40 out of 45 targets. This was a 66% relative improvement over the OneMap baseline, which often failed due to misidentifying objects when spatial relationships were not considered.
The researchers note that while the system performs well, challenges remain with small or hidden objects and occasional LVLM validation failures. Future work aims to address these limitations and generalize to even more complex spatial reasoning beyond proximity. For more details, you can read the full research paper: DIV-Nav: Open-Vocabulary Spatial Relationships for Multi-Object Navigation.
