StopWatch tackles a quiet failure in object-goal navigation: vision-language-model (VLM) robots often reach the requested object but keep walking because they cannot tell they have arrived. Its two-model design separates navigation from goal grounding—the decision that the scene matches the requested goal and the robot should stop—while corrected evaluations reveal stronger results and show where perception still falls short.
What Did the Researchers Build?
StopWatch is a system for object-goal navigation: a robot receives a target such as an object category and must move through an environment until it reaches a suitable example. The task has two parts. The robot must navigate to the target, and it must recognize when it has arrived and stop. StopWatch makes that second capability—goal grounding—an explicit part of the system rather than leaving it buried inside one end-to-end navigation policy.
The system uses two models in a pipeline. One handles navigation; the other supports grounding, connecting what the robot sees with the requested goal and the decision to stop. The grounding capability can be trained from scene annotations and improved separately from the navigation policy. That separation lets researchers investigate whether a failure comes from moving through the environment or from recognizing that the destination has been reached.
The work also addresses how navigation systems are tested. The researchers report that official evaluations can reject valid stops, distorting comparisons between agents. They provide a corrected evaluation and introduce two open-vocabulary benchmarks, MP3D-OV and HSSD-OV, extending testing beyond HM3D.
What Results Did StopWatch Achieve?
The central finding is that reaching the goal and succeeding at the task are not the same thing. On novel goal categories, an end-to-end policy comes within 0.25 metres of the target in 76% of episodes, but its recorded success rate is only 52%. The researchers identify failure to recognize arrival—and therefore failure to stop—as the main reason for much of this gap.
| Reported measure | Result |
|---|---|
| Episodes coming within 0.25 m of a novel-category goal | 76% |
| Episodes counted as successful | 52% |
| Difference between the two rates | 24 percentage points |
The 24-point difference is calculated from the two reported rates; it is not a separate benchmark result. It highlights why judging navigation only by proximity can tell a different story from judging whether the robot completed the task.
The researchers report that StopWatch reaches state-of-the-art results across all the object-goal navigation tasks they benchmarked. They also find that existing official evaluations reject valid stops, and reject more of them for stronger agents. The paper’s conclusion identifies the corrected evaluation and the new MP3D-OV and HSSD-OV benchmarks as resources for more faithful comparisons. The supplied results do not include per-benchmark scores, so the size of StopWatch’s advantage cannot be quantified here.
How Does StopWatch Work?
A conventional end-to-end navigation policy must use visual input and a goal description to choose actions, while also learning when to end the episode. If the robot gets close to the right object but continues moving, the failure can look like a navigation problem—even if the robot has already done the difficult work of finding the target. StopWatch treats the arrival decision as a distinct capability that deserves its own supervision and evaluation.
In broad terms, the pipeline assigns navigation and grounding to separate models. The navigation model is responsible for moving through the scene toward the goal. The grounding component addresses whether the observed scene supports the requested goal and whether the robot has reached a valid stopping point. Scene annotations provide supervision for this grounding task, allowing it to be improved independently rather than relying only on action demonstrations to teach the whole policy what to do.
The paper’s analysis also changes how success should be interpreted. A proximity measure asks whether the robot came near the goal; a task-success measure asks whether it completed the episode correctly, including stopping. Those measures can diverge substantially. A benchmark that rejects legitimate stop actions adds another source of error, making an agent appear less capable than it is. StopWatch therefore pairs a system-level change with a cleaner evaluation approach and new open-vocabulary test sets.
Why Does This Matter for Robotics?
A mobile robot that reaches the right place but fails to stop has not completed its job. In a warehouse or facility, that could mean passing a requested item, continuing into an occupied area, or making a delivery interaction harder to verify. Separating “move toward the goal” from “confirm the goal and stop” gives engineers a clearer way to diagnose these failures and decide which capability needs improvement.
The lesson also applies to how robot systems are selected and compared. A single success score can hide whether a system struggles to find targets, recognize them, or end an episode at the right time. Better evaluation makes those differences visible, which matters when a team compares navigation approaches or checks whether a system is ready for a specific environment. See warehouse robots on Robot Overflow for examples of robots used in facility operations.
The new open-vocabulary benchmarks broaden the testing picture beyond HM3D. That matters because a navigation policy can look capable on familiar categories yet struggle with less familiar goals. StopWatch’s results point to a practical development priority: once navigation is strong, reliable visual understanding of the target and arrival condition can be the larger obstacle.
What Are the Limitations and Open Questions?
The reported results establish a benchmark finding, not proof of dependable operation in physical workplaces. The paper’s supplied results do not establish how the system performs with changing lighting, moving people, blocked views, or objects that are difficult to distinguish. Those conditions can complicate both recognizing a target and deciding when to stop.
The reported 76% and 52% figures describe one setting involving novel goal categories; they do not show how the gap changes across environments or target types. The paper reports state-of-the-art performance but the available text does not provide per-benchmark scores, making the size of the improvement difficult to assess. The corrected evaluation also means results from older and revised scoring procedures may not be directly comparable. Further testing should show whether the separate grounding component remains reliable across varied scenes and real robot deployments.
What Questions Do Readers Ask About StopWatch?
What is goal grounding in robot navigation? Goal grounding is the process of connecting a requested goal to what the robot sees and determining whether it has reached an appropriate stopping point.
Why can a robot reach an object and still fail? A robot can get near the target but fail to recognize that it has arrived, so it walks past instead of stopping.
What does StopWatch change? It separates navigation from goal grounding in a two-model pipeline and addresses flaws in the evaluation used to score navigation.
What are MP3D-OV and HSSD-OV? They are open-vocabulary object-goal navigation benchmarks introduced by the researchers to extend evaluation beyond HM3D.
What Is the Main Takeaway?
StopWatch shows that navigation is not only about reaching the right place; it is also about recognizing arrival and stopping correctly. Its separate grounding component and corrected benchmarks make that often-hidden part of robot performance easier to study.