Many AI technologies depend on accurately reconstructing dynamic 3D environments, from self-driving vehicles to immersive virtual reality. However, representing the wide variety of motions found in real-world scenes remains a fundamental challenge because no single dynamic representation can consistently model the diverse motions encountered in practice. Different Dynamic Gaussian Splatting (DGS) methods perform well under specific conditions, as each motion representation has its own strengths and limitations. As a result, a single representation often struggles to generalize across heterogeneous real-world dynamics.
To address this limitation, a team of researchers led by Professor Kyeongbo Kong from Pusan National University developed two complementary Mixture-of-Experts frameworks that combine the strengths of multiple motion representations. MoE-GS independently trains multiple dynamic Gaussian models and then adaptively blends their outputs through learned expert routing. In contrast, MoDE integrates multiple deformation experts during joint optimization using a shared Gaussian representation. Rather than relying on a single motion model, these frameworks leverage the complementary strengths of multiple specialized models to improve dynamic scene reconstruction. This paper was made available online on July 13, 2026 in the IEEE Transactions on Pattern Analysis and Machine Intelligence.
"We conducted a systematic analysis, and have come to the understanding that no existing Dynamic Gaussian Splatting method consistently performs best across diverse scenarios. Motivated by this finding, we introduce MoE-GS, the first framework that adaptively combines multiple specialized dynamic Gaussian models through a Mixture-of-Experts architecture instead of relying on a single representation," explains Prof. Kong.
The researchers found that combining complementary experts enable more accurate reconstruction of complex scenes containing multiple types of motion than relying on a single dynamic representation. The adaptive routing strategy allows the framework to blend the most appropriate expert for each region and time step, improving reconstruction quality while maintaining flexibility across diverse dynamic scenarios.
The team believes this adaptive approach could enhance AI systems used in robotics, digital twins, autonomous systems, and spatial computing by enabling more reliable reconstruction of complex environments with heterogeneous motion. Looking ahead, they suggest that accurately modeling diverse real-world dynamics will become increasingly important as AI evolves toward interacting more naturally with the physical world. They hope their work will provide a valuable foundation for future research in dynamic scene understanding, world models, and Physical AI.
Finally, Prof. Kong concludes, "Our findings suggest that combining multiple specialized motion representations can be an effective way to handle heterogeneous dynamics that are difficult for a single representation to model consistently."
Reference
Title of original paper: On the Design of Mixture-of-Experts for Dynamic Gaussian Splatting
Journal: IEEE Transactions on Pattern Analysis and Machine Intelligence
DOI: https://doi.org/10.1109/tpami.2026.3712736
About Pusan National University
Pusan National University, located in Busan, South Korea, was founded in 1946 and is now the No. 1 national university of South Korea in research and educational competency. The multi-campus university also has other smaller campuses in Yangsan, Miryang, and Ami. The university prides itself on the principles of truth, freedom, and service and has approximately 30,000 students, 1,200 professors, and 750 faculty members. The university comprises 14 colleges (schools) and one independent division, with 103 departments in all.
Website: https://www.pusan.ac.kr/eng/Main.do
About the author
Prof. Kyeongbo Kong is an Associate Professor in the Department of Electrical & Electronics Engineering at Pusan National University, where he leads the Computer Vision & Signal Processing (CVSP) Laboratory. His research aims to develop scalable and reliable Spatial AI systems that can understand and model dynamic real-world environments through advances in computer vision, multimodal AI, and 3D/4D world modeling. His group develops AI technologies for dynamic scene reconstruction, robot perception, and vision-language models. He received his M.S. and Ph.D. in Electrical Engineering from POSTECH.
Lab website: https://www.pnu-cvsp.com/
ORCID: 0000-0002-1135-7502
