EviNav: Evidence-Driven Grounding and Validated Termination for Zero-Shot Instance Navigation
EviNav is a training-free, evidence-driven framework for zero-shot instance navigation that accumulates multi-view visual and semantic evidence to robustly distinguish targets and determine termination, achieving state-of-the-art success rates in both image- and text-goal navigation tasks.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a robot sent into a room it has never seen before, tasked with finding a specific object. The challenge is not merely to find a chair or a table, but to locate one particular chair or table that matches a specific description or a single photograph provided by a human. This is the core of instance navigation. Unlike simpler tasks where a robot just needs to find any object of a certain type, this requires the machine to distinguish between nearly identical items, navigate around obstacles, and decide exactly when it has found the right one. The difficulty lies in the fact that the robot cannot simply guess; it must gather enough visual proof to be certain, all while managing a strict limit on how many steps it can take before the task ends. If it stops too early, it might pick the wrong object; if it keeps searching too long, it runs out of steps and fails.
Researchers at the University of Sussex have developed a new system called EviNav to solve this problem without needing to train the robot on specific rooms or objects beforehand. Instead of learning from thousands of examples, the system uses a set of logical rules and powerful, pre-existing artificial intelligence tools to reason about what it sees in real time. The system works by building a mental map of the environment as it moves, tracking potential targets, and constantly asking itself: "Is this the right object, and do I have enough evidence to stop?" The researchers tested this approach in thousands of simulated scenarios and on a physical robot in real offices and laboratories, finding that it significantly outperforms previous methods by being more careful about how it gathers proof and when it decides to finish.
The heart of the system is a method for accumulating evidence rather than making snap judgments. When the robot is given a text description, such as "a white toilet with a yellowing seat," it does not simply stop at the first white toilet it sees. Instead, it treats every candidate object as a hypothesis that needs verification. The robot moves to different angles around the object, taking pictures from distinct viewpoints. It then uses a vision-language model—a type of artificial intelligence that can understand both images and text—to check if the object matches the description. Crucially, the system understands that taking multiple pictures from the exact same spot does not provide new information. It only counts evidence gathered from different positions. If the robot is unsure, it actively moves to a new angle to get a better look, rather than just asking the same question again. This process continues until the accumulated proof is strong enough to confirm the identity of the object or until the robot is forced to move on.
For tasks where the goal is a photograph, the system takes a different but equally rigorous approach. A photograph of an object implies a specific camera position and angle from which that photo was taken. The researchers found that simply moving toward the object shown in the picture is often not enough, because the robot might end up standing behind a wall or in a spot where the object is obscured. Instead, EviNav analyzes the features in the reference photo and the live video feed to calculate exactly where the original camera must have been standing. It then navigates to that specific location and aligns its own camera to match the original angle before checking if the object is indeed the target. This ensures the robot sees the object in the same context as the reference image, avoiding false matches caused by similar-looking furniture or architectural details.
A major innovation in this work is how the system manages its limited time. The robot operates under a strict budget of steps; it cannot wander forever. The system keeps track of the best candidate it has found so far and reserves enough steps to return to that candidate if it runs out of time to find a better one. This "fallback" mechanism prevents the robot from getting lost in a search that goes on too long. If the robot spends too many steps exploring and fails to confirm a target, it automatically turns back to the most promising spot it visited earlier, ensuring it at least has a chance to stop at a likely candidate before time runs out. The researchers also implemented a way to filter out areas that look like floors but are actually elevated surfaces, like tables or beds, ensuring the robot only plans paths across walkable ground.
The results of these experiments were substantial. In simulated tests involving one thousand different episodes for both text-based and image-based goals, the system succeeded in finding the correct object 84.1 percent of the time for image goals and 55.0 percent for text goals. These numbers represent a significant improvement over previous methods, which often struggled with false positives or ran out of steps. The system was also tested on a real robot equipped with a stereo camera in an office and a laboratory. In these physical trials, the robot succeeded in 20 out of 24 attempts, successfully navigating real-world clutter, sensor noise, and uncalibrated smartphone photos used as reference images. The researchers noted that the system achieved these high success rates while using far less computing power and making fewer requests to external artificial intelligence services than other leading methods.
The study explicitly rules out the idea that simply getting closer to an object or matching its surface features is sufficient for success. In their tests, methods that navigated directly to the surface of a matched object without verifying the camera angle or gathering evidence from multiple viewpoints failed much more often. The researchers also demonstrated that relying on a single view to make a decision leads to errors, as repeated checks from the same spot can reinforce a wrong conclusion. By contrast, their system's requirement for distinct viewpoints and its ability to recover the original camera position proved to be the primary drivers of its success. The work suggests that for robots to navigate complex, unseen environments reliably, they must be designed to treat stopping as a deliberate verification process, gathering independent evidence from different perspectives before committing to a final decision.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.