\nVisualization: An indoor space is captured by a mobile robot. Objects that were visible only moments earlier and are now hidden behind furniture or outside the field of view are shown grayed out and unlabeled on the left half of the image. On the right half, the same objects are shown color-segmented, carrying stable ID labels, and connected by fine glowing lines to a central “Memory Bank” symbol | Image: \u00a9 Ulrich Buckenlei | XR Stager Online Magazine | VISORIC GmbH
\n<\/sup><\/p>\n
<\/p>\n
Computer vision was long considered a solved problem, as long as objects moved within the camera frame. Classic multi-object trackers link positions from frame to frame, as long as motion stays predictable and an object remains visible. But the moment a robot turns a corner, an object disappears behind a piece of furniture, or lighting and viewing angle change, that chain breaks. For the system, the world starts over with every re-encounter.<\/p>\n
A research team at Universidad Polit\u00e9cnica de Madrid set out to address exactly this problem. With REMIND, the scientists present a fully open-source system that re-identifies objects not through motion, but through visual identity, even after long interruptions, from different viewing angles, and under changing lighting conditions.[1]<\/sup><\/p>\n This shifts a fundamental assumption in robotics. Until now the rule was: a system must continuously observe an object to know it. Persistent object recognition reverses this principle. An object remains part of a system’s memory even if it was out of sight for minutes, as long as its visual signature can be recognized again.[2]<\/sup><\/p>\n For robotics, digital twins, and physical AI, this is more than a technical footnote. Persistent object memory is a basic requirement for autonomous systems to work reliably in real, cluttered environments, without starting from zero every time they return to a room.<\/p>\n The real innovation, however, does not lie in a single technique, but in the interplay of modern visual foundation models, a two-stage memory store, and a built-in mechanism that would rather admit uncertainty than guess wrong. Only this interplay turns another tracking algorithm into a system that truly remembers.<\/p>\n This article shows how persistent object recognition works, why it is becoming a central building block of physical AI, and what applications it enables in robotics, logistics, building technology, and digital twins. It contextualizes the research from CVAR-UPM, explains the underlying technology, and uses current examples to show why machines without memory eventually hit their limits.<\/p>\n <\/p>\n Multi-object tracking was originally developed for video analysis: people and vehicles are meant to be followed from frame to frame, usually at high frame rates and with largely continuous visibility. For surveillance cameras or traffic analysis, this principle is sufficient. For mobile robots moving freely through indoor spaces, it is not.<\/p>\n A service robot moving down a corridor constantly loses objects from view. A tool cart disappears behind a door, a box is hidden by a shelf, a machine sits outside the field of view for several minutes. Classic trackers interpret every one of these interruptions as a loss. When the object reappears later, it is assigned a new, independent identity, even though it is the same physical object.<\/p>\n REMIND addresses exactly this weak point. Instead of extending motion trajectories, the system builds a visual representation of each object that stays stable regardless of camera position, lighting, and viewing angle. When an object reappears after an interruption, the system compares its current appearance with the stored representations and restores the original identity instead of creating a new one.<\/p>\n This approach is made possible by modern visual foundation models. REMIND uses frozen DINOv3 features as the basis for object description. These features were not trained for a single tracking scenario; instead, they learn general visual representations from vast amounts of image data that are robust to changes in viewing angle and lighting. This robustness is exactly what matters for re-identifying objects across time and space.[3]<\/sup><\/p>\n <\/p>\n The difference sounds technical at first, but it has immediate practical consequences. A warehouse robot that recounts the same pallet after every turn corrupts inventory data. An assistance system that registers a tool as a new object after every occlusion cannot maintain a meaningful history. Persistent memory fixes this fundamental problem by decoupling identity from continuous visibility.<\/p>\n This also changes the requirements for the underlying system architecture. It is no longer enough to track positions over a few frames. A system needs a memory that keeps object identities stable across minutes, rooms, and changing conditions, without creating exploding compute or memory demands.<\/p>\n For a memory store like this to actually work, a single stored image per object is not enough. How REMIND balances reliability and flexibility, and why visual identity becomes the true foundation of robotic perception, is the subject of the next chapter.<\/p>\n <\/p>\n Photographing an object once is not enough to reliably recognize it later. Lighting changes, the viewing angle shifts, parts of the object get occluded, and several similar objects often stand next to each other in the same room. Five identical office chairs, several identical cardboard boxes, or a row of similar machine components pose a considerable challenge for a simple image-matching system.<\/p>\n REMIND addresses this with a two-stage memory architecture. Newly detected objects first enter a “Working Bank”: a provisional representation that is still uncertain. Only once an object is recognized repeatedly and consistently does its representation move into a “Stable Bank” with higher confidence. This principle resembles the way humans initially place new faces tentatively and only recognize them with certainty after several encounters.[4]<\/sup><\/p>\n A second crucial building block is looking at object parts rather than treating whole objects as a single, indivisible image patch. Instead of treating a chair as one uniform feature, REMIND considers the backrest, seat, and legs separately. This keeps recognition stable even when an object is partially occluded or only visible from an unusual angle.<\/p>\n Equally important is the context of the surrounding environment. Five identical chairs around the same table can hardly be distinguished by appearance alone. REMIND therefore incorporates neighboring objects as additional context for identification: which objects are nearby, in what spatial arrangement, at what relative position to one another. This contextual information significantly increases discriminative power, especially for near-identical objects.<\/p>\n <\/p>\n A final, often underestimated building block is the deliberate handling of uncertainty. Instead of forcing an identity in every case, REMIND can explicitly flag an object as “ambiguous” when the available evidence is insufficient. For an assistance system that makes decisions based on recognized objects, this distinction is crucial: a wrong identity can trigger a wrong decision, while a correctly flagged doubt instead leads to an additional check.<\/p>\n These four principles, tiered memory, part-based representation, spatial context, and controlled uncertainty, together form a system that is significantly more robust than simple image comparison. On the research team’s own indoor benchmark, REMIND achieves an IDF1 score of 90.35 percent, roughly 20 points ahead of the strongest video segmentation method and more than 36 points ahead of a strong tracking-by-detection approach.[5]<\/sup><\/p>\n A robust object memory is the foundation. Its real value, however, only unfolds once it becomes part of a larger spatial information model that connects robotics, buildings, and digital twins. That is exactly the subject of the next chapter.<\/p>\n <\/p>\n Digital twins represent real environments as spatial models and connect them with real-time data. Until now, this was mostly based on a snapshot: a scan, a camera, or a sensor delivers the current state of a scene, but as soon as objects move or become occluded, the model quickly loses accuracy. Persistent object recognition fundamentally changes this principle by giving a digital twin a continuous memory for moving objects.<\/p>\n Instead of a static image, a model emerges that knows where a particular tool was last located, which transport cart belongs to which order, or which shelf was most recently restocked, even if no sensor is capturing the object at that exact moment. An object’s identity persists in the system even when its visibility is interrupted.<\/p>\n For warehouses, production environments, and buildings with high object density, this is a significant advance. A digital twin of a warehouse no longer has to reinterpret from scratch, with every camera switch, which object is where. Instead, it maintains a continuous object history assembled from multiple cameras, points in time, and viewing angles, without identities being lost or duplicated.<\/p>\n The basis for this is the benchmark evaluation on ScanNet++, a dataset for high-resolution indoor reconstructions. REMIND leads in nearly every tested configuration. Particularly notable: the system completes 100 percent of the test scenes, while a comparison method fails on two-thirds of the scenes due to memory overflow. For practical, resource-constrained robotics and digital-twin applications, this stability matters at least as much as raw accuracy.[6]<\/sup><\/p>\n <\/p>\n This merges two previously separate disciplines: the static geometry of classic digital twins and the dynamic, continuously updating object layer of robotics. A building model no longer shows only walls, machines, and fixed installations, but also mobile inventory objects with their own history, state, and identity over time.<\/p>\n For operators of complex environments, this opens up new ways of tracking, without any additional RFID tags or manual data entry. Persistent visual object recognition complements existing sensor systems with a layer based purely on camera images, which makes it particularly cost-effective to integrate into existing infrastructure.<\/p>\n A digital twin with object memory provides the data foundation. How artificial intelligence uses this information to independently recognize what it has already seen, and when it should be unsure of its own assessment, is the subject of the next chapter.<\/p>\n <\/p>\n A system that only recognizes objects without maintaining a history remains limited in what it can tell us. The real value of persistent object recognition only emerges once artificial intelligence actively uses this continuous history to draw connections, assess states, and support decisions.<\/p>\n For an assistance system in a workshop, this means: a tool is not merely recognized, but its position is tracked across the last work shift. For a service robot system in a hospital, it means a piece of medical equipment can be found again even if it was moved to a different room in the meantime. Object history itself becomes a source of information, not just a technical side condition.<\/p>\n What matters here is how REMIND handles uncertainty. Instead of forcing a binary decision at every re-identification, the system continuously assesses how confident an assignment is. If the evidence is insufficient, an object is explicitly flagged as uncertain rather than being assigned a false identity. For downstream AI systems, such as planning algorithms or language assistants, this distinction is crucial: information flagged as uncertain can be specifically verified, while a silently wrong assignment corrupts every decision built on top of it.<\/p>\n <\/p>\n This ability for calibrated self-assessment is what separates a research-mature system from a mere demonstration technology. In safety-relevant applications, such as intralogistics, healthcare, or industrial maintenance, the ability to communicate “I’m not sure” is often more important than a marginally higher average accuracy.<\/p>\n Beyond that, the continuously updated object history can be linked with further AI systems. Language models can use the stored object identities to answer natural-language questions, such as where a particular tool was last seen or how a warehouse’s state has changed over the past hour. Pure object recognition thereby becomes a queryable, context-aware knowledge base.<\/p>\n A system that reliably remembers and correctly assesses its own uncertainty is relevant to numerous industries, well beyond pure research scenarios. Which concrete application areas already benefit from persistent object recognition today is the subject of the next chapter.<\/p>\n <\/p>\n The real value of persistent object recognition is not shown in a single demonstration, but in the breadth of its application fields. Anywhere objects move through unstructured, changing environments and repeatedly disappear from view, a visual long-term memory delivers direct, practical benefit.<\/p>\n In warehouse automation, persistent object recognition enables more reliable inventory tracking without any additional sensors. Mobile robots navigating shelves, aisles, and picking areas can consistently identify objects across multiple passes instead of recounting them at every re-encounter.<\/p>\n In indoor navigation, service robots benefit from fixed and moving landmarks being recognized stably over time. A cleaning robot, a delivery robot, or an assistance system in an office building can not only orient itself spatially, but also understand which objects have changed since the last pass.<\/p>\n In healthcare and care facilities, persistent object recognition can help locate medical devices, aids, or supply carts without attaching any additional tracking hardware. In building technology, it helps facility management systems keep track of mobile inventory such as tools, cleaning equipment, or temporary fixtures.<\/p>\n <\/p>\n New possibilities also emerge within production itself. Tools, fixtures, and semi-finished products can be tracked consistently across multiple workstations, without every station needing its own dedicated capture system. For quality assurance and traceability, this continuous object identity is a considerable advantage over isolated, station-specific capture systems.<\/p>\n What all these application fields have in common is a fundamental property: they require no specialized per-object setup and no additional physical marking. Object identity results purely from visual appearance, which makes the system comparatively easy to integrate into existing camera infrastructure.<\/p>\n As different as these application fields are, they share a common principle: humans and machines increasingly draw on the same continuous object history. How this changes collaboration between humans and robots is the subject of the next chapter.<\/p>\n <\/p>\n Until now, a robot’s knowledge of its environment mostly stayed confined to the system itself. A person entering the same room had no direct access to what the robot had previously observed. With persistent object recognition, a shared, queryable object history emerges for the first time, one that both machines and humans can access.<\/p>\n A service technician entering a facility can directly see which objects have changed over the past few hours, without having to search every corner themselves. A warehouse manager can query where a particular object was last registered, regardless of which robot or camera last captured it. Object history thereby becomes a shared knowledge base that is not tied to a single device.<\/p>\n This effect is especially clear in human-robot teams working together in the same environment. When a robot hands off a task to a human colleague, for example during picking or maintenance, it can immediately communicate which objects are relevant, where they are located, and how confident that information is. Conversely, a human can give the system a correction as soon as an object was incorrectly flagged as uncertain, thereby specifically improving system confidence.<\/p>\n <\/p>\n This form of shared perception significantly reduces friction. Instead of passing on information verbally or capturing it again, everyone involved works from the same up-to-date state. Misjudgments can be corrected faster because uncertainties are made transparent rather than staying hidden inside a closed system.<\/p>\n In the long run, this also changes the role of robots in mixed teams. They become less of an isolated tool and more of a participant in a shared information space, in which observations, object identities, and confidence levels are continuously exchanged between humans and machines.<\/p>\n A shared knowledge base like this only works if the underlying technology remains open, understandable, and extensible. Why exactly this is a defining feature of REMIND is the subject of the next chapter.<\/p>\n <\/p>\n Many high-performing perception systems in robotics are developed within individual companies and remain proprietary. For research, further development, and industry-wide standardization, this is a significant obstacle: results are hard to verify independently, further development remains limited to individual vendors, and smaller research teams or companies have no access to comparably capable systems.<\/p>\n CVAR-UPM takes a different approach with REMIND. Code, evaluation framework, and training data have been made fully publicly available.[7]<\/sup> Developers and research teams can not only use the system, but understand its inner workings in detail, integrate it into their own robotics and computer vision projects, and extend it in targeted ways.<\/p>\n This open approach considerably accelerates progress across the entire research field. Other teams can build on the same foundations instead of developing comparable systems from scratch. Results can be meaningfully compared against other methods, because benchmark data and evaluation methodology are openly available. For a research field that depends heavily on reproducible comparisons, this transparency is of central importance.<\/p>\n <\/p>\n For companies outside pure research, this openness means a lower barrier to entry into a demanding technology. Instead of developing a comparable system entirely from scratch, REMIND can be used as a foundation for one’s own robotics, logistics, or building applications and adapted specifically to the given environment.<\/p>\n At the same time, the open nature of the project shows how progress in physical AI is increasingly taking hold: not through individual closed platforms, but through openly accessible research that is jointly advanced by a broad community across science and industry.<\/p>\n Open research is the foundation. Its true significance, however, only unfolds in the broader context of a development that reaches far beyond a single research project. Why persistent object recognition is becoming a central building block of the next generation of physical AI is the subject of the final chapter.<\/p>\n <\/p>\n Physical AI describes systems that do not merely process language or images, but perceive, understand, and act within the physical world. Robots, autonomous vehicles, and intelligent building systems belong here, as do assistance systems that continuously interpret real objects and environments. For these systems to work reliably, they need more than momentary recognition. They need memory.<\/p>\n Until now, much research has focused on perceiving the current moment: which object is visible right now, which action is happening right now, which gesture is being performed right now. Persistent object recognition extends this focus by adding a temporal dimension. A system understands not only what is happening now, but also what happened before, and can keep this information consistent across minutes, rooms, and changing conditions.<\/p>\n This extension is not an isolated advance, but fits into a larger developmental trajectory. Digital twins are becoming more dynamic, robotics is becoming more context-aware, and artificial intelligence increasingly gains access to structured, reliable information about the physical world, rather than merely to momentary sensor snapshots.<\/p>\n <\/p>\n For companies investing in robotics, automation, or digital twins, this has a clear consequence. Systems that only recognize objects at the moment of their visibility will eventually hit their limits once they need to handle real, cluttered environments with changing conditions. Systems with persistent memory, by contrast, are built from the ground up for exactly these conditions.<\/p>\n Research projects such as REMIND show that this capability is already technically feasible, scientifically grounded, and openly accessible today. The path from an academic publication to broad industrial application has become significantly shorter than it was just a few years ago.<\/p>\n This brings the article full circle. What begins with a simple problem, robots forgetting objects the moment they leave the field of view, develops into a foundational technology for the next generation of robotics, digital twins, and physical AI. How this principle shows up in a concrete application is illustrated in the following video.<\/p>\n <\/p>\n <\/p>\n The previous chapters have shown how persistent object recognition works technically, why it is becoming a foundational building block of physical AI, and which application fields, from robotics to building technology, benefit from it. This principle becomes especially vivid, however, once you watch such a system at work directly.<\/p>\n The following video shows raw footage from a typical residential and office environment, as published as part of CVAR-UPM’s REMIND research. Objects such as remote controls, laptops, chairs, plants, and cushions are segmented live, given stable ID numbers, and consistently re-identified across multiple camera angles and points in time. Colored outlines mark each detected object, and fine lines connect objects to their respective labels.<\/p>\n The observation over time is particularly revealing: objects that are briefly occluded, for example by a passing person or a camera movement, retain their original label as soon as they become visible again. This behavior is exactly what distinguishes persistent object recognition from classic, short-lived tracking.<\/p>\n\n
From Tracking to Memory<\/h2>\n
![\"Split](\"https:\/\/www.xrstager.com\/wp-content\/uploads\/2026\/06\/visoric-physical-ai-object-memory-001.jpg\")
<\/p>\nClassic tracking loses an object as soon as it leaves the field of view and assigns a new identity when it reappears. Persistent object memory recognizes the same object by its visual signature and automatically restores its original identity.<\/h6>\n
\nInfographic: Comparison of classic short-term tracking and persistent object recognition with visual long-term memory | Graphic: \u00a9 Ulrich Buckenlei | XR Stager Online Magazine | VISORIC GmbH
\n<\/sup><\/p>\n\n
Why Visual Identity Becomes the New Foundation<\/h2>\n
![\"Technical](\"https:\/\/www.xrstager.com\/wp-content\/uploads\/2026\/06\/visoric-physical-ai-object-memory-002.jpg\")
<\/p>\nPersistent object recognition combines a two-stage memory store, part-based features, and the spatial context of neighboring objects. Only this combination enables reliable re-identification even among near-identical objects.<\/h6>\n
\nInfographic: Architectural principle of persistent object recognition with Working Bank, Stable Bank, part-based features, and environmental context | Graphic: \u00a9 Ulrich Buckenlei | XR Stager Online Magazine | VISORIC GmbH
\n<\/sup><\/p>\n\n
Digital Twins Gain a Long-Term Memory<\/h2>\n
![\"Interior](\"https:\/\/www.xrstager.com\/wp-content\/uploads\/2026\/06\/visoric-physical-ai-object-memory-003.jpg\")
<\/p>\nA digital twin with persistent object memory maintains a continuous history of moving objects, even when they are temporarily outside the visibility range of cameras or sensors.<\/h6>\n
\nInfographic: Persistent object memory as an extension of classic digital twin models for warehouse, production, and building environments | Graphic: \u00a9 Ulrich Buckenlei | XR Stager Online Magazine | VISORIC GmbH
\n<\/sup><\/p>\n\n
Artificial Intelligence Recognizes What It Has Seen Before<\/h2>\n
![\"Close-up](\"https:\/\/www.xrstager.com\/wp-content\/uploads\/2026\/06\/visoric-physical-ai-object-memory-004.jpg\")
<\/p>\nAn AI-powered assistance system evaluates not only which object was recognized, but also how confident that assignment is. Uncertain cases are specifically flagged rather than silently treated as correct.<\/h6>\n
\nInfographic: Confidence-based object recognition as the foundation of trustworthy AI assistance systems in robotics and building technology | Graphic: \u00a9 Ulrich Buckenlei | XR Stager Online Magazine | VISORIC GmbH
\n<\/sup><\/p>\n\n
A Technology for Robotics, Logistics and Buildings<\/h2>\n
![\"Six-part](\"https:\/\/www.xrstager.com\/wp-content\/uploads\/2026\/06\/visoric-physical-ai-object-memory-005jpg.jpg\")
<\/p>\nPersistent object recognition applies across industries: from warehouse automation and indoor navigation to healthcare and facility management, through to production and field service.<\/h6>\n
\nInfographic: Application fields of persistent object recognition in robotics, logistics, building technology, and industrial maintenance | Graphic: \u00a9 Ulrich Buckenlei | XR Stager Online Magazine | VISORIC GmbH
\n<\/sup><\/p>\n\n
Humans and Machines Remember Together<\/h2>\n
![\"Factory](\"https:\/\/www.xrstager.com\/wp-content\/uploads\/2026\/06\/visoric-physical-ai-object-memory-006.jpg\")
<\/p>\nRobots and humans access the same persistent object history. Observations, identities, and uncertainties are shared in real time instead of remaining isolated within a single system.<\/h6>\n
\nInfographic: Shared object history as the foundation for collaborative human-robot work environments | Graphic: \u00a9 Ulrich Buckenlei | XR Stager Online Magazine | VISORIC GmbH
\n<\/sup><\/p>\n\n
Open Research Instead of Closed Systems<\/h2>\n
![\"Circular](\"https:\/\/www.xrstager.com\/wp-content\/uploads\/2026\/06\/visoric-physical-ai-object-memory-007.jpg\")
<\/p>\nFully open access to code, research results, and training data allows research institutions, startups, and industrial companies alike to build on the same technological foundation.<\/h6>\n
\nInfographic: Open access to research, code, and datasets as the foundation for broad further development of persistent object recognition | Graphic: \u00a9 Ulrich Buckenlei | XR Stager Online Magazine | VISORIC GmbH
\n<\/sup><\/p>\n\n
Physical AI as the Next Evolutionary Stage of Robotics<\/h2>\n
![\"Central](\"https:\/\/www.xrstager.com\/wp-content\/uploads\/2026\/06\/visoric-physical-ai-object-memory-008.jpg\")
<\/p>\nPersistent object recognition, together with visual foundation models, contextual understanding, and calibrated uncertainty, forms the technological foundation for the next generation of physical AI.<\/h6>\n
\nInfographic: Persistent object recognition as a foundational technology for robotics, digital twins, logistics, building technology, and human-machine collaboration within physical AI | Graphic: \u00a9 Ulrich Buckenlei | XR Stager Online Magazine | VISORIC GmbH
\n<\/sup><\/p>\n\n
When a Robot Truly Remembers<\/h2>\n