Autonomous 3D active mapping couples reconstruction with action selection: at each step, a robot must update an incomplete map and choose the observation that will improve it next. This closed loop is especially consequential for planetary surface or subsurface reconnaissance, where supervision is delayed, motion is costly, and reliable external localization may be unavailable. Learned geometric completion is attractive in this setting. An occupancy model can use partial observations to predict unobserved structure and guide sensing before a complete map exists.
Introduction
The same prior creates a coupled risk. In recent active-mapping systems, one occupancy map contributes to the expected new surface visible from a candidate viewpoint and also constrains collision-free motion. A hallucinated surface may therefore attract the robot as an apparently informative target, block a route through free space, or do both. When the mapper must operate between communication windows, planner-facing geometry must remain revisable as onboard evidence accumulates.
We therefore ask: when should an active mapper retain a learned geometric prior, and when should accumulated observations override it? We use a controlled terrestrial 3D benchmark to isolate this planner-facing mechanism. The experiments do not claim validation under planetary appearance, degraded sensing, or long-horizon localization drift. They instead diagnose the shared occupancy interface and test an online revision rule for the closed perception, mapping, and planning loop.
Related Work
We focus on this perception–planning interface. Rather than changing the occupancy architecture, we freeze a planner, intervene on its geometry, and trace changes in viewpoint utility and coverage. The proposed online correction layer preserves predictions in sparsely observed space and suppresses those that remain unsupported after repeated nominal exposure. This interface matters when learned completion guides action during long periods of onboard autonomy.
Planner-Facing Occupancy
Let the occupied set exposed to the planner and the candidate trajectories considered by the planner define the shared planning interface. The interface combines rendered predicted surface gain with collision-feasibility testing.
A single occupied prediction can therefore change the objective, the feasible set, or both. A false positive may look like an uncovered target and may also reject an otherwise traversable route; a false negative may remove a useful target or change inferred connectivity.
These effects unfold in closed loop. The selected trajectory determines the next RGB-D measurements, those measurements update the planner-facing occupancy representation and the learned prediction, and the new planner-facing geometry changes the next planning problem. Pointwise occupancy accuracy at one step cannot fully characterize this downstream effect. Our interventions measure the net result after the feedback has unfolded.
Findings
First, offline ground-truth tags reveal that phantom occupancy contributes 53.5% and 35.0% of the rendered gain in two viewpoints selected by the learned planner. The error therefore enters the decision signal, rather than remaining only a map-quality defect.
This non-monotonicity is consistent with sequential exploration. A local correction can reorder candidate trajectories, after which the robot observes a different subset of the scene. It may accelerate early coverage without improving the final reachable surface, or remove a misleading target that had incidentally carried the robot toward useful nearby geometry. The gap between ground-truth improvements in AUC and final coverage further indicates that candidate-view placement, reachability, and trajectory search remain limiting even after geometry is corrected. World models therefore require decision-level evaluation alongside geometric accuracy.
Conclusion
Our results motivate rethinking learned occupancy in autonomous active mapping as a revisable planning prior. Because learned occupancy enters both gain estimation and collision checking, its errors can alter where an active mapper looks and which motions it considers feasible. Controlled interventions show that occupancy accuracy alone does not predict closed-loop coverage. Observation-gated filtering retains useful completion while suppressing repeatedly unsupported predictions, improving two targeted failure cases without retraining or ground truth.
Although evaluated here for 3D active mapping, keeping planner-facing predictions revisable also matters in active search. Search-TTA updates its image encoder during search with uncertainty-weighted online gradients, refining potentially inaccurate target-presence maps. Its parameter adaptation differs from our deterministic, reversible 3D occupancy gate, yet both use accumulated task evidence to revise predictions before subsequent planning decisions. We consider active search as a future work.
Frequently Asked Questions
What problem does learned occupancy create in autonomous active mapping? Hallucinated surfaces can attract the robot toward misleading targets, block free-space routes, or do both.
What does observation-gated filtering do? It preserves predictions in sparsely observed space and suppresses predictions that remain unsupported after repeated nominal exposure.
Does occupancy accuracy alone predict closed-loop coverage? No. Candidate-view placement, reachability, trajectory search, and feedback effects also influence final coverage.
Was the method evaluated in planetary environments? No. The experiments use a controlled terrestrial 3D benchmark and do not claim validation under planetary appearance, degraded sensing, or long-horizon localization drift.