UniSim-SLAM Keeps Robots on Track in Unfamiliar Spaces

Abstract Recent geometric foundation models enable feed-forward inference for SLAM, but their predictions are strongly dependent on the input view set, which leads to geometric inconsistencies and trajectory drift when results are chained over long sequences. Online deployment further exposes a trade-off between the low latency of two-view tracking and the constraint richness of multi-view inference. We introduce UniSim-SLAM, an integrated system that runs lightweight two-view keyframe tracking in the frontend and performs periodic multi-view submap refinement in the backend. To combine predictions defined in heterogeneous local coordinates with inconsistent scales, we formulate a unified multi-level factor graph on Sim(3) that jointly optimizes global keyframe poses and submap poses. The graph integrates temporal view-to-view odometry edges, view-to-submap bridge edges with depth-statistics scale anchoring, and submap-to-submap tie and scale constraints to enforce consistent similarity relations across submaps. Experiments on TUM RGB-D and 7-Scenes show that UniSim-SLAM achieves state-of-the-art accuracy in the uncalibrated setting, reducing trajectory error by 38.5% on TUM RGB-D and 45.9% on 7-Scenes compared to prior best results. Robots navigating unfamiliar environments must track their position while continuously mapping the space around them. AI has made both tasks faster, but a persistent trade-off remains: using more camera views can improve accuracy, but requires more processing time. Professor Kyungdon Joo and his research team from the Graduate School of Artificial Intelligence at UNIST have developed UniSim-SLAM, a system that combines rapid camera tracking with periodic multi-view refinement. The approach allows robots to update their position quickly while correcting errors that would otherwise accumulate as they move. Known as simultaneous localization and mapping (SLAM), the underlying technology enables robots and other autonomous systems to determine their location while reconstructing their surroundings. Recent AI-based approaches can infer 3D geometry directly from camera images, but small differences between successive estimates can compound over long sequences, gradually pushing the reconstructed path away from the robot's actual trajectory. UniSim-SLAM tackles the problem by giving fast tracking and detailed reconstruction different roles. It compares two camera views at a time to track movement, then periodically brings multiple views together to refine the local 3D map. A unified optimization framework aligns estimates produced at different scales and orientations, keeping the reconstructed environment and camera trajectory consistent over time. On two standard indoor benchmarks achieved, UniSim-SLAM state-of-the-art results in the uncalibrated setting. With the previous best Compared methods, it reduced trajectory error by 38.5% on TUM RGB-D and 45.9% on 7-Scenes. The system was also considerably faster than a multi-view alternative evaluated in the study. VGGT-SLAM required 3.41 seconds to produce a result, compared with 0.197 seconds for UniSim-SLAM. The new approach also improved the accuracy of 3D reconstruction. “UniSim-SLAM brings fast tracking and accurate multi-view analysis into the same system,” said Professor Joo. “This could benefit applications that need reliable localization and real-time 3D reconstruction, from robotics and autonomous driving to augmented reality.” The study, with Inha Lee of UNIST as first author, was presented at the European Conference on Computer Vision (ECCV 2026), held September 8–12 in Malmö, Sweden. The work was supported by the Ministry of Science and ICT (MSIT) through the Institute of Information & Communications Technology Planning & Evaluation (IITP), including the Artificial Intelligence Graduate School Program at UNIST and programs in collaborative AI and spatial immersive technologies, as well as the InnoCORE program. Journal Reference Inha Lee, Dongjae Jeong, Junhee Lee, and Kyungdon Joo, "UniSim-SLAM: Feed-Forward SLAM with Unified Sim(3) Optimization," ECCV'26 (2026).

2026.10.08

A More Efficient Energy Strategy for Power-Hungry AI Data Centers

AI data centers consume enormous amounts of electricity—not only to run thousands of GPUs, but also to keep them cool. A new study suggests that recovering waste heat while integrating on-site power generation and carbon capture could substantially reduce both their cost and carbon footprint. Led by Professor Hankwon Lim of the Graduate School of Carbon Neutrality at UNIST, a research team evaluated a data-center energy system combining solid oxide fuel cells (SOFCs), carbon capture, and organic Rankine cycle (ORC) cooling. Compared with a conventional grid-powered, air-cooled system, the proposed approach could reduce total system costs by approximately 48% and lifecycle greenhouse gas emissions by up to 72%. The system puts waste heat to work in two ways. High-temperature heat from the fuel cells supplies energy for the carbon-capture process, while heat removed from servers is recovered through an ORC to generate additional electricity. This reduces both the energy required for cooling and the amount of electricity the data center must draw from the grid. The researchers modeled a data center with 20,000 GPUs and compared eight combinations of power supply, cooling, and carbon capture using energy prices and electricity mixes in Korea, the United States, and the European Union. Across a broad range of cost and emissions priorities, the combination of SOFC power, carbon capture, and ORC cooling consistently ranked highest. ORC cooling performed particularly well across all three regions, reducing annual cooling-related electricity costs by as much as $19 million. Lower electricity demand also cuts annual greenhouse gas emissions by an estimated 80,000 to 150,000 metric tons, potentially avoiding $5 million to $10 million in annual carbon costs under carbon-pricing schemes such as the EU Emissions Trading System. The preferred source of electricity varied by region. Relatively inexpensive natural gas made SOFC generation economically attractive in the United States, while in Korea, where natural gas is more expensive, grid electricity remained more economical when cost was prioritized. “As AI data centers expand, electricity supply, cooling, and carbon emissions need to be considered together,” said Donghyoun Lee from the Graduate School of Carbon Neutrality at UNIST. “Our analysis shows how power generation, waste-heat recovery, and carbon capture can be evaluated as a single energy system.” “The best strategy will depend on local energy prices and the carbon intensity of electricity,” said Professor Lim. “This framework can help identify energy systems that balance cost and emissions for data centers of different sizes and in different regions.” The findings of this study have been published online in Chemical Engineering Journal on August 14, 2026. The research was supported by the National Research Foundation of Korea (NRF) and the Ministry of Science and ICT (MSIT). Journal Reference Donghyoun Lee, JeHyeon Seong, and Hankwon Lim, "An integrated solution for data center power supply, thermal management, and carbon capture evaluated through a multi-criteria decision analysis framework," Chem. Eng. J., (2026).

2026.10.07

EPIC Helps Personal AI Remember What Matters

Abstract With the rapid emergence of personal AI agents based on Large Language Models (LLMs), implementing them on-device has become essential for privacy and responsiveness. To handle the inherently personal and context-dependent nature of real-world requests, such agents must ground their generation in device-resident personal context. However, under tight memory budgets, the core bottleneck is what to store so that retrieval remains aligned with the user. We propose EPIC (Efficient Preference-aligned Index Construction), which focuses on user preferences as a compact and stable form of personal context and integrates them throughout the RAG pipeline. EPIC selectively retains preference-relevant information from raw data and aligns retrieval toward preference-aligned contexts. Across four benchmarks covering conversations, debates, explanations, and recommendations, EPIC reduces indexing memory by 2,404 times, improves preference-following accuracy by 18.79 %p, and achieves 32.17 times lower retrieval latency over the best-performing baseline. In on-device experiments, EPIC maintains under 1 MB memory and achieves 5.21 to 29.35 ms/query latency across three platforms, while supporting streaming updates under preference drift. From previous conversations to individual preferences and constraints, personal AI agents work best when they can draw on a user's own context. However, keeping all of that information readily searchable can quickly strain the limited memory of smartphones and other personal devices. Led by Professor Taesik Gong of the Department of Computer Science and Engineering at UNIST, the team developed EPIC (Efficient Preference-aligned Index Construction), an on-device retrieval framework that selectively stores information relevant to a user's preferences. The approach sharply reduces memory use and retrieval time while improving how well AI responses reflect individual needs. EPIC is designed for retrieval-augmented generation (RAG), which allows a language model to draw on information beyond what it learned during training. On personal devices, however, indexing everything from documents and browsing activity to conversation histories can consume substantial memory. EPIC instead filters data before indexing, retaining what matters to the user while discarding what does not. For each item it retains, EPIC also creates an instruction describing when that information should be used. If a user with a seafood allergy asks for food recommendations while traveling, for example, the system can connect the request with that dietary constraint and retrieve suitable options rather than simply returning the most popular dishes. Across four benchmarks covering recommendations, debates, explanations, and conversations, EPIC reduced indexing memory by up to 2,404 times and retrieval latency by 32.17 times compared with the best-performing baseline. With Llama-3.1-8B-Instruct, preference-following accuracy improved by an average of 18.79 percentage points. Tests on a Jetson Orin Nano, MacBook Pro M4, and Galaxy Z Flip6 showed that EPIC could operate with less than 1 MB of retrieval memory, retrieving relevant information in 5.21 to 29.35 milliseconds per query. The system can also update its memory as new information arrives or user preferences change, without rebuilding the entire index. “EPIC is well suited to personalized AI that must adapt as new information arrives and user preferences change,” said Changmin Lee, the study's first author. “We plan to explore more flexible retrieval strategies, including ways to retain useful information that may be only indirectly related to a preference, while removing or compressing outdated memories.” “Rather than storing as much information as possible, this work focuses on storing what is valuable to the individual user,” said Professor Gong. “EPIC could help make personalized AI faster and more practical on everyday devices while keeping personal information on the device.” The research was presented at the 43rd International Conference on Machine Learning (ICML 2026) in Seoul. The code and data are available through the team's EPIC repository. The work was supported by the Ministry of Science and ICT and the Institute of Information & Communications Technology Planning & Evaluation (IITP), as well as the National Research Foundation of Korea. Journal Reference Changmin Lee, Jaemin Kim, and Taesik Gong, "From Volume to Value: Preference-Aligned Memory Construction for On-Device RAG," ICML '26 , (2026).

2026.10.06

Central Pacific El Niño Improves North Atlantic Winter Forecasts

Abstract The wintertime North Atlantic Oscillation is a major mode of variability affecting climate across the Northern Hemisphere, yet its seasonal prediction skill over the North Atlantic remains lower than that over the Pacific. We examine North Atlantic Oscillation predictability through Pacific teleconnections using reanalysis data and multi-model hindcasts from seasonal-to-decadal forecasting systems. Here we show that Central Pacific El Niño events significantly enhance both North Atlantic Oscillation forecast skill and the predictable signal over the North Atlantic. Reanalysis and equatorial Pacific pacemaker experiments reveal that Central Pacific El Niño enhances sea surface temperature variability over the Aleutian region, modulates Atlantic storm-track activity, and strengthens the Pacific–Atlantic teleconnection. Coupled Model Intercomparison Project Phase 6 simulations suggest that more frequent Central Pacific El Niño occurrence under greenhouse warming may favor this teleconnection. These results identify Central Pacific El Niño-driven Pacific–Atlantic teleconnections as an important source of North Atlantic Oscillation predictability and seasonal forecast skill. Seasonal forecasts for the North Atlantic have long lagged behind those for the Pacific. A new study finds that forecasts of the North Atlantic Oscillation (NAO) become considerably more accurate during winters marked by Central Pacific El Niño. A team, led by Professor Myong-In Lee of the Department of Civil, Urban, Earth, and Environmental Engineering at UNIST traced the improvement to a stronger climate connection between the North Pacific and North Atlantic. The NAO is a major pattern of atmospheric variability that shapes winter weather across Europe and North America. The study involved researchers from the Korea Institute of Science and Technology (KIST) and the Met Office Hadley Center in the United Kingdom. Across several seasonal forecasting systems, the correlation between predicted and observed NAO conditions for January and February was 0.48 over the full study period. During Central Pacific El Niño winters, it rose to 0.78, showing a marked increase in forecast skill. The researchers found that Central Pacific El Niño strengthens a chain of climate interactions linking the two ocean basins. It increases sea surface temperature variability near the Aleutian region of the North Pacific, shifting the position and intensity of the storm track. The resulting changes in atmospheric circulation extend across North America to the North Atlantic, where they influence the NAO. Because ocean temperatures change more slowly than the atmosphere, these North Pacific temperature patterns can persist long enough to provide an early signal of atmospheric conditions that develop later over the North Atlantic. “The North Atlantic has strong internal atmospheric variability, which makes predictable signals difficult for climate models to capture,” said Kiwook Kim, the study's first author. “During Central Pacific El Niño, the stronger connection between the Pacific and Atlantic effectively opens a window of opportunity for more skillful NAO forecasts The findings also shed light on the signal-to-noise paradox, in which the real atmosphere proves more predictable than climate models suggest. The researchers found that models may underestimate the connection between North Pacific sea surface temperatures and North Atlantic circulation, limiting their ability to capture this predictable signa. “Where El Niño develops in the Pacific matters for how its influence reaches the North Atlantic,” said Professor Lee. “If climate models can better capture this connection, we may be able to improve seasonal outlooks for the NAO and related winter extremes, including cold spells and heavy snowfall. The study was published in Communications Earth & Environment , a sister journal to Nature Portfolio journals, on August 27, 2026. The research was supported by the Korea Meteorological Administration and by the Ministry of Science and ICT (MSIT) through the National Research Foundation of Korea (NRF). Journal Reference Kiwook Kim, Myong-In Lee, Daehyun Kang, and Adam A. Scaife, “Central Pacific El Niño-driven Pacific–Atlantic teleconnections are an important source of North Atlantic Oscillation predictability,” Commun. Earth Environment. , (2026).

2026.10.02

Beyond Ideal Conditions

《Editor's Note: Research works by controlling variables. But the problems it ultimately confronts rarely offer the same courtesy. Heat gets into storage tanks. Expertise shifts from one question to the next. Satellite images arrive out of season. Materials face conditions far beyond the laboratory norm. The obvious response is to restore control —keep the heat out, remove the variation, recreate the ideal test. But what happens when ideal conditions are no longer an option? This month, UNIST researchers took that question in different directions. Some intervene before a problem takes hold —or after a familiar line of defense has failed. Others design for rather variation than uniformity, or put promising results to more demanding tests. The approaches differ, but the underlying shift is the same —rather than trying to restore control, they rethink the solution around what actually changes.》 CHANGE WHERE YOU INTERVENE “Sometimes the best place to solve a problem is not where it appears.” For atomically thin semiconductors, contamination during fabrication can be especially difficult to undo. Plasma exposure changes the chemistry of photoresist, causing residue to bind more strongly to MoS₂. Aggressive cleaning, meanwhile, risks damaging the material itself. Rather than trying to remove the residue afterward, UNIST researchers placed a removable, 5-nanometer metal layer over the semiconductor before patterning, preventing the contamination from reaching its surface. The resulting transistors showed roughly a tenfold reduction in contact resistance and a similar increase in on-current. Liquid hydrogen presents the problem from the opposite direction. Even well-insulated tanks cannot keep heat out indefinitely. Instead of treating insulation as the only line of defense, researchers investigated whether metal-organic frameworks could capture evaporating hydrogen after heat enters the tank and slow the buildup of pressure. In modeling, IRMOF-20 preserved about 97 percent of the volumetric capacity of liquid hydrogen while extending the projected storage period from roughly 64 to 221 days. One solution acts before the problem reaches the material: the other takes effect after a familiar line of defense has been breached. In both cases, changing where researchers intervene changes what can be done about the problem. ■ New Process Shields 2D Semiconductors from Fabrication Residue (Small) ■ Rethinking Liquid Hydrogen Storage with MOFs (Nature Communications) WORK WITH WHAT CHANGES “When the world refuses to stay consistent, the method has to change with it.” Personalization gets harder when expertise refuses to stay fixed. A person who is an expert in one subject may be a novice in another—and even within the same field, expertise can change from question to question. ExPerT responds to that variation rather than smoothing it away. By combining the semantics of each query with patterns in a user's typing behavior, the framework estimates expertise for individual questions and adjusts the detail, terminology, and complexity of its answers accordingly. In a study involving 40 participants and 1,270 queries, it reduced expertise-estimation error by 65.7 percent compared with the strongest baseline. For wildfire assessment, the changing conditions are not the user but the landscape. Satellite images taken before and after a fire may come from different seasons, allowing natural changes in vegetation to be mistaken for fire damage. Instead of requiring seasonally matched images, UNIST researchers use nearby unburned vegetation to estimate seasonal change and correct the comparison. Across 12 US wildfires, it achieved a correlation of 0.72 with field observations using seasonally mismatched imagery—nearly matching the 0.73 obtained with images from similar seasons. One method adapts to knowledge that changes from question to question—the other accounts for landscapes that change with the season. Neither requires the variation to disappear before the analysis can begin. ■ AI Personalization Adapts Answers to What Users Know (ACL '26) ■ Assessing Wildfire Burn Severity Across Seasons (Remote Sens. Environ.) TEST WHAT HOLDS UP “A promising result means more when it holds up under a more demanding test.” Perovskite solar cells (PSCs) have reached impressive efficiencies, but performance under controlled conditions tells only part of the story. UNIST researchers developed a molecular interface that delivered a certified power-conversion efficiency of 26.37 percent, then subjected the cells to repeated swings between −100°C and +100°C. After 16 cycles, they retained about 91 percent of their initial efficiency—putting their stability to a far more demanding test. AI highlight detection faces a different gap between promising results and meaningful tests. Existing benchmarks have largely relied on videos lasting only a few minutes, even though real sports broadcasts can run for hours. SVHighlights closes the gap with 320 videos across eight sports and 640.18 hours of footage. Its videos average about two hours—roughly 30 to 60 times longer than those in existing datasets—allowing AI models to be evaluated under conditions much closer to the task they are meant to perform. One study subjects a material to extreme temperature swings, while the other tests AI against the scale of the task it is meant to perform. Both ask the same question: Does a promising result still hold when the test starts to resemble the challenge it is meant to meet? ■ PSCs Hold Up Under Extreme Temperature Swings (Joule) ■ AI Sports Highlight Detection to a More Realistic Test (ACM KDD '26) Explore the latest issue and follow 『The UNIST Brief』 on LinkedIn to stay updated. Subscribe via LinkedIn

2026.10.01

Rethinking Liquid Hydrogen Storage with Metal-Organic Frameworks

Liquid hydrogen (LH₂) is attractive for long-distance energy transport because of its high volumetric energy density. But even well-insulated tanks cannot completely prevent heat from entering. As the liquid warms, hydrogen evaporates and pressure builds, leading to boil-off losses during storage and transport. A research team, led by Professor Hyunchul Oh of the Department of Chemistry at UNIST, together with Professor Hoi Ri Moon of Ewha Womans University, and researchers at the Technical University of Munich in Germany, has investigated whether porous materials could help limit these losses. The team used metal-organic frameworks (MOFs)—highly porous crystalline materials—to capture evaporating hydrogen and delay its release as heat enters the tank. MOFs contain networks of nanoscale pores that can absorb hydrogen onto their internal surfaces. At cryogenic temperatures, interactions between hydrogen molecules and the pore walls help keep the gas confined, slowing pressure buildup. The approach complements conventional insulation—instead of focusing only on limiting heat entering the tank, it also changes how hydrogen responds once heat gets in. The key question is whether that benefit comes at the expense of storage capacity. A porous material takes up space that would otherwise hold liquid hydrogen. The researchers therefore compared two MOFs with very different structures—IRMOF-20, a rigid framework with a large pore volume, and MIL-53(AI), a flexible framework whose pores expand and contract as hydrogen is adsorbed. IRMOF-20 offers a particularly favorable balance. When the researchers accounted for the space occupied by the material, the system retained about 97% of the volumetric capacity of neat liquid hydrogen. This was possible because hydrogen confined within the pores packed densely enough to compensate for much of the space taken by the MOF. The researchers then modeled how the materials would affect boil-off in a transport-scale liquid hydrogen tank. Under the mid-vacuum insulation conditions examine, neat liquid hydrogen was predicted to deplete after about 64 days. With IRMOF-20, that period increased to approximately 221 days—more than three times as long. Experiments offered clues to why IRMOF-20 performed this way. Adsorption measurements indicated that hydrogen confined within its pores reached an effective density higher than that of bulk liquid hydrogen. Inelastic neutron scattering also showed restricted molecular rotation inside the pores, providing indirect evidence of strong interactions between hydrogen and the framework. Together, the results suggest that nanoscale confinement can both pack hydrogen dedensely and stabilizze it against release as temeprature rises. MIL-53(AI) behaved differently. It retained hydrogen more strongly as temperature increased, but its volumetric capacity was only about 53% of that of liquid hydrogen. The contrast highlights a central design choice for porous materials in cryogenic hydrogen storage—larger pore volume favors capacity, while stronger confinement can improve thermal retention. “By considering pore volume, hydrogen density, and desorption behavior together, we found that poros materials could reduce boil-off while preserving much of the storage capacity needed for liquid hydrogen transport,” said Professor Oh. “These calculations represent an idealized upper bound, so further work is needed to determine how closely this performance can be reproduced in practical tank systems.” Jaewoo Park of UNIST served as first author, with contributions from Dr. Monica Jiménez-Ruiz of the Institut Laue-Langevin in France and Dr. Margarita Russina of Helmholtz-Zentrum Berlin fur Materialien und Energie in Germany, among others. The research was supported by the National Research Lab (NRL 2.0) Program, including the Institute for Multiscale Matter and Systems (IMMS), and the BrainLink program of the National Research Foundation of Korea. The findings were published in Nature Communications on September 7, 2026.

2026.09.30