Scene graphs that store evidence, not decisions.

Maps that make up their mind too early can't take it back

A sofa seen from four viewpoints may be proposed as sofa, chair, bench, or armchair — occlusion, scale, and foundation-model noise guarantee disagreement. ConceptGraphs, HOV-SG, FSR-VLN, and DualMap all resolve that disagreement at construction time, exposing each object as one fused feature, label, or consensus state. Related approaches push on adjacent axes without changing this pattern: ConceptFusion and VLMaps embed features densely rather than per object, and OGScene3D and ThinkGraphs refine tracking and confidence while still exposing a single committed state per object. Three things go wrong.

Four stages, each pushing commitment later

An object node stores the evidence-preserving state \(\mathcal{B}_i = (P_i, \mathbf{f}_i^v, \mathcal{O}_i, \mathbf{V}_i, \mathbf{p}_i, \mathbf{f}_i^t, \rho_i)\) — point cloud, aggregate visual feature, the raw language observations, concept-level evidence with its smoothed belief, a consensus text feature, and a trust status. The visual and language channels are never destructively fused during mapping; every cross-modal decision is deferred to a readout.

Overview of OpenBelief-Nav. Mapping retains observation-level phrases, reliability cues, frame–mask provenance, and separate aggregate geometric and visual representations. Task-time readouts perform fixed-vocabulary projection or free-form retrieval, while arrival verification enables query-specific candidate correction without modifying the persistent graph.

Watch a belief being built — and vote splitting break it

Five views of the same sofa arrive with different phrases. Play the stream, then toggle synonym consolidation off: the lexical variants sofa / couch / loveseat split their votes and a single high-support chair observation wins. Drag α to see Dirichlet smoothing keep sparse evidence from collapsing into false certainty.

“What is this object?” depends on who's asking

Because a node stores evidence rather than a decision, the same frozen graph answers differently — and correctly — for each task.

Navigation is a physical test of the hypothesis

A retrieval result is a hypothesis, not a fact. On arrival, the robot's RGB-D observation runs an ordered verifier — re-observe if the candidate isn't sufficiently visible (absence of view is not evidence of absence), reject on semantic mismatch or on a semantic match that fails to coincide geometrically with the stored node, confirm otherwise. Rejection changes exactly one thing: a per-query exclusion set \(\mathcal{R}_q^{(k+1)} = \mathcal{R}_q^{(k)} \cup \{i^{(k)}\}\). The graph is never modified and the query is never rewritten.

Same graph, strongest result on both benchmarks

Graph construction is performed once per environment; every number below is a readout over that frozen graph, with no re-perception between tasks.

Deployed on a Unitree G1 humanoid

Compute-heavy semantic reasoning runs offboard; an onboard Jetson keeps the motion loop local. The bridge is Zenoh. The robot navigates to a candidate, looks, and lets physics vote.

Real-world platform and deployment procedure. Left: the Unitree G1 with RealSense D435i and Livox MID-360. Right: execution-grounded candidate correction — arrival observations confirm, re-observe, or reject the selected candidate, and rejection re-grounds the unchanged query over the remaining mapped candidates without modifying the persistent graph.

BibTeX

@preprint{2026openbeliefnav,
  title   = {OpenBelief-Nav: Evidence-Preserving Open-Vocabulary Scene Graphs
             for Reliable Language-Guided Robot Navigation},
  
  author  = {anoynimous},
  year    = {2026},
  note    = {Under review}
}