A new memory architecture for indoor robots treats each camera frame as transient evidence and fuses it into a three level semantic map that holds across revisits, instead of storing it as a permanent state.
Indoor robots can navigate one room. Making them remember an entire building, across hallways, floors, and revisits to rooms they mapped weeks ago, is the hard part. A new paper, GEM-Occ, proposes a memory architecture that treats every camera frame as transient evidence to be fused, not stored.
GEM-Occ organizes a robot's spatial memory into three nested layers: a local cache for the agent's current view, a room-level submap for the space it occupies, and a building-level graph that stitches those submaps together. Any layer can be queried at any time through Gaussian-to-occupancy splatting, a way of converting a probabilistic 3D representation into a queryable semantic occupancy grid. That hierarchy is what lets an agent answer queries like locating the kitchen relative to the elevator without re-scanning the building.
The central move is to treat each frame's visual geometry as evidence, not as a persistent state. The model converts a local geometry prediction into two kinds of evidence: semantic Gaussian occupancy (where something is) and free-space ray evidence (where something is not). A visibility- and uncertainty-aware causal update folds that evidence into the matching submap. The system absorbs small per-frame noise without storing the noise itself, which is why revisit consistency can improve without explicitly re-observing prior geometry.
The paper also ships HIOcc, a benchmark that unifies ScanNet, ScanNet++, and Matterport3D, three existing indoor scene datasets, under one sparse semantic-occupancy format. It supports three evaluation regimes: local semantic occupancy prediction in a single frame, room-level online mapping as the agent moves, and building-level mapping across connected panoramic environments. Reported gains over baselines show up across local occupancy prediction, online map stability, free-space reasoning, revisit consistency, and building-level scalability.
Indoor embodied-agent research is in the middle of crossing the single-room boundary. SLAM and map-fusion pipelines that work in a single room tend to drift across corridor changes, accumulate stale geometry on revisits, and break down when the agent returns to rooms mapped weeks earlier. GEM-Occ enters that conversation as a concrete architectural choice, with one benchmark for other teams to compare against.
The contribution is incremental, not a breakthrough, and the paper has not yet cleared peer review. Reported improvements are over baselines on the authors' own unified benchmark, not against deployed indoor-robot systems, and not against human-level spatial reasoning. "Semantic occupancy" remains a research construct; there is no field deployment, no real-robot rollout, and no claim of robustness in outdoor or mixed indoor-outdoor settings. The benchmark itself is a reformatting of three existing datasets, not new data.
The code and project page are public, so the architecture will be testable by other indoor-robot teams before any peer-review status lands. Watch items: whether HIOcc becomes a shared yardstick across the indoor-occupancy community, whether the Gaussian-evidence trick survives independent reproduction outside the authors' code release, and whether the three-level memory hierarchy generalizes beyond the three indoor datasets the benchmark currently uses.