Event

DGIST's LT-Mem Wins Best Paper at IROS 2026

1 reports · 1 sources

The paper LT-Mem from the APRL Lab at the Daegu Gyeongbuk Institute of Science and Technology (DGIST) in South Korea won the Best Paper award at IROS 2026. The work proposes a volatility-aware spatiotemporal memory framework that uses three layers of memory—Live, Delta, and Meta—and three update strategies based on volatility—coverage, retention, and multi-hypothesis—to address the temporal forgetting problem in which new observations overwrite old memories when a robot revisits the same space multiple times. According to the report, LT-Mem is a memory-layer extension to existing robot perception systems; its operation depends on front-end modules such as multi-session SLAM, instance segmentation, and cross-session object identity matching. The experimental hardware was an RTX 5070 Ti GPU and 64GB of memory, and it is better suited to conventional indoor and semi-outdoor scenes where static objects dominate and dynamic objects are a minority. The report also notes that the STAR method's QA-Event metric uses keyword-matching scoring, whereas other methods use strict-matching scoring, so differences in scoring criteria should be noted when making cross-method comparisons.