{"id":56165,"date":"2026-08-27T14:53:30","date_gmt":"2026-08-27T13:53:30","guid":{"rendered":"https:\/\/www.intelligentcio.com\/apac\/?p=56165"},"modified":"2026-08-27T14:53:30","modified_gmt":"2026-08-27T13:53:30","slug":"south-koreas-pusan-national-university-develops-multi-expert-ai-framework-for-dynamic-3d-reconstruction","status":"publish","type":"post","link":"https:\/\/www.intelligentcio.com\/apac\/2026\/08\/27\/south-koreas-pusan-national-university-develops-multi-expert-ai-framework-for-dynamic-3d-reconstruction\/","title":{"rendered":"South Korea\u2019s Pusan National University develops multi-expert AI framework for dynamic 3D reconstruction"},"content":{"rendered":"\n

Researchers at South Korea\u2019s Pusan National University have developed a new approach to dynamic 3D reconstruction that could help AI systems build more accurate representations of complex, constantly changing real-world environments.<\/em><\/p>\n\n\n\n

Research from South Korea\u2019s Pusan National University addresses one of the persistent problems facing technologies ranging from autonomous vehicles and robotics to digital twins, virtual reality and spatial computing: how to reconstruct a three-dimensional scene when different objects are moving in different ways.<\/p>\n\n\n\n

Rather than attempting to find a single model capable of handling every type of movement, the research team has taken a multi-model approach. Led by Professor Kyeongbo Kong, the researchers developed two complementary Mixture-of-Experts (MoE) frameworks designed to combine specialised motion representations according to the characteristics of a scene.<\/p>\n\n\n\n

The work could have implications for the development of Physical AI and world models, where machines need increasingly sophisticated ways to perceive, understand and interact with dynamic physical environments.<\/p>\n\n\n\n

No single model fits every scene<\/strong><\/p>\n\n\n\n

Dynamic 3D reconstruction has become increasingly important as AI moves beyond static images and towards systems capable of understanding environments that change over time.<\/p>\n\n\n\n

Dynamic Gaussian Splatting (DGS) has emerged as a promising technique for reconstructing such scenes. However, individual DGS approaches tend to excel under particular conditions because different representations have different strengths and weaknesses.<\/p>\n\n\n\n

A technique capable of accurately modelling one form of movement may perform less effectively when confronted with another. Real-world environments can contain multiple forms of motion simultaneously, making the reliance on one representation a significant limitation.<\/p>\n\n\n\n

The researchers conducted a systematic analysis of existing approaches and concluded that no single DGS method consistently delivered the strongest results across diverse scenarios.<\/p>\n\n\n\n

That finding led the team to investigate whether several specialised models could instead work together.<\/p>\n\n\n\n

“We conducted a systematic analysis, and have come to the understanding that no existing Dynamic Gaussian Splatting method consistently performs best across diverse scenarios,” said Kong.<\/p>\n\n\n\n

“Motivated by this finding, we introduce MoE-GS, the first framework that adaptively combines multiple specialised dynamic Gaussian models through a Mixture-of-Experts architecture instead of relying on a single representation.”<\/p>\n\n\n\n

Bringing in the experts<\/strong><\/p>\n\n\n\n

At the centre of the work are two approaches, MoE-GS and MoDE, which apply the Mixture-of-Experts principle in different ways.<\/p>\n\n\n\n

Mixture-of-Experts architectures divide a problem among specialised models or ‘experts’, allowing different parts of a system to concentrate on particular characteristics of the data rather than requiring one model to handle everything.<\/p>\n\n\n\n

MoE-GS independently trains multiple dynamic Gaussian models before adaptively combining their outputs through learned expert routing.<\/p>\n\n\n\n

The routing mechanism determines which experts are most appropriate for particular regions and time steps, allowing the system to draw on different representations as conditions within a dynamic scene change.<\/p>\n\n\n\n

MoDE takes a different route. Instead of separately training multiple complete dynamic Gaussian models, it integrates multiple deformation experts during joint optimisation while using a shared Gaussian representation.<\/p>\n\n\n\n

The distinction enables the researchers to investigate different trade-offs involving model flexibility, reconstruction performance and computational efficiency.<\/p>\n\n\n\n

Importantly, both approaches move away from the assumption that a single representation should be responsible for modelling every movement within a complex scene.<\/p>\n\n\n\n

Tackling heterogeneous motion<\/strong><\/p>\n\n\n\n

The researchers found that combining complementary experts enabled more accurate reconstruction of scenes containing multiple types of motion compared with relying on a single dynamic representation.<\/p>\n\n\n\n

Adaptive routing is particularly significant because a complex environment is unlikely to require the same modelling technique everywhere.<\/p>\n\n\n\n

A scene encountered by an autonomous machine, for example, might simultaneously contain moving people, vehicles, objects and changing backgrounds. The nature, direction and complexity of those movements can differ substantially.<\/p>\n\n\n\n

By learning how to blend the most appropriate experts for different spatial regions and points in time, the framework can adapt its representation to the motion it encounters.<\/p>\n\n\n\n

That capability could become increasingly valuable as AI systems are expected to understand less controlled environments.<\/p>\n\n\n\n

The research also illustrates a broader direction within AI development: rather than attempting to build increasingly complex universal models for every individual problem, systems can potentially combine specialised components and dynamically select the expertise required for a particular task.<\/p>\n\n\n\n

From digital twins to Physical AI<\/strong><\/p>\n\n\n\n

More reliable reconstruction of dynamic environments could have applications across several rapidly developing areas of technology.<\/p>\n\n\n\n

Robotics is an obvious candidate. Machines operating alongside people need to understand not only where objects are located but how their surroundings are changing and how those changes are likely to develop.<\/p>\n\n\n\n

Autonomous vehicles face a similar requirement as they interpret environments containing pedestrians, cyclists, other vehicles and moving infrastructure.<\/p>\n\n\n\n

Digital twins could also benefit. As digital representations become more sophisticated, accurately capturing movement and change in the physical assets or environments they replicate will become increasingly important.<\/p>\n\n\n\n

Spatial computing and immersive virtual reality present another potential application, particularly where virtual representations need to respond accurately to movement occurring in physical space.<\/p>\n\n\n\n

The work may ultimately have wider significance for Physical AI; an area focused on AI systems capable of perceiving and interacting with the physical world.<\/p>\n\n\n\n

Such systems will require more than conventional recognition capabilities. They will need representations that can accommodate complex environments in which multiple objects behave differently and conditions continuously evolve.<\/p>\n\n\n\n

World models face much the same challenge. If AI is to develop internal representations of physical environments that can be used to predict events and guide actions, accurately understanding heterogeneous motion will be a fundamental requirement.<\/p>\n\n\n\n

The research suggests that combining specialised representations could provide one route towards solving that problem.<\/p>\n\n\n\n

“Our findings suggest that combining multiple specialised motion representations can be an effective way to handle heterogeneous dynamics that are difficult for a single representation to model consistently,” Kong said.<\/p>\n\n\n\n

Rather than searching for one dynamic representation capable of mastering every environment, the research points towards a more adaptive future in which multiple experts work together. As AI increasingly moves from analysing digital information to understanding and acting within the physical world, that flexibility could prove increasingly important.<\/p>\n","protected":false},"excerpt":{"rendered":"

Researchers at South Korea\u2019s Pusan National University have developed a new approach to dynamic 3D reconstruction that could help AI systems build more accurate representations of complex, constantly changing real-world environments. Research from South Korea\u2019s Pusan National University addresses one of the persistent problems facing technologies ranging from autonomous vehicles and robotics to digital twins, […]<\/p>\n","protected":false},"author":58,"featured_media":56166,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[38,8983,44,11274],"tags":[852,5787,2662,12056,12055,12059,12054,12060,11429,12057,7132,6009,5401,12058],"class_list":["post-56165","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-analysis-research","category-c-suite-technology","category-top-stories","category-university","tag-artificial-intelligence","tag-autonomous-vehicles","tag-digital-twins","tag-dynamic-3d-reconstruction","tag-dynamic-gaussian-splatting","tag-mixture-of-experts-2","tag-mode","tag-moe-gs","tag-physical-ai","tag-pusan-national-university","tag-robotics","tag-spatial-computing","tag-virtual-reality","tag-world-models"],"acf":[],"publishpress_future_workflow_manual_trigger":{"enabledWorkflows":[]},"_links":{"self":[{"href":"https:\/\/www.intelligentcio.com\/apac\/wp-json\/wp\/v2\/posts\/56165","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.intelligentcio.com\/apac\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.intelligentcio.com\/apac\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.intelligentcio.com\/apac\/wp-json\/wp\/v2\/users\/58"}],"replies":[{"embeddable":true,"href":"https:\/\/www.intelligentcio.com\/apac\/wp-json\/wp\/v2\/comments?post=56165"}],"version-history":[{"count":1,"href":"https:\/\/www.intelligentcio.com\/apac\/wp-json\/wp\/v2\/posts\/56165\/revisions"}],"predecessor-version":[{"id":56167,"href":"https:\/\/www.intelligentcio.com\/apac\/wp-json\/wp\/v2\/posts\/56165\/revisions\/56167"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.intelligentcio.com\/apac\/wp-json\/wp\/v2\/media\/56166"}],"wp:attachment":[{"href":"https:\/\/www.intelligentcio.com\/apac\/wp-json\/wp\/v2\/media?parent=56165"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.intelligentcio.com\/apac\/wp-json\/wp\/v2\/categories?post=56165"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.intelligentcio.com\/apac\/wp-json\/wp\/v2\/tags?post=56165"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}