Event
GEMS Spatiotemporal Memory Framework and Benchmark Released
1 reports · 1 sources
Ling-Chia Chen and other authors propose the GEMS framework, which combines a persistent spatiotemporal 3D Gaussian Splatting (3DGS) map with an agent reasoning pipeline for long-term spatiotemporal memory. The authors state that the map is reconstructed only once, after which only regions where the scene changes are updated and those changes are saved, in order to preserve visual detail over time; with the help of the map, the agent renders new viewpoints on demand, progressively accumulates evidence to answer spatiotemporal questions, and localizes objects according to natural language queries. The authors also release the GEMS-Bench benchmark, comprising 3 real environments and 12 synthetic environments, with a total of 473 question-answer pairs.
Report timeline
- GEMS: A Long-Term Spatiotemporal Memory Framework Combining 3D Gaussian Splatting Maps with Agentic Reasoning
The authors propose the GEMS framework, which combines a persistent spatiotemporal 3D Gaussian Splatting (3DGS) map with an agentic reasoning pipeline. The map is reconstructed once and thereafter updated only in regions where the scene changes, with those modifications saved. Using this map, the agentic pipeline renders novel views on demand, progressively accumulates evidence to answer spatiotemporal questions, and grounds objects from natural-language queries.
arXiv cs.RO ·