Most spatial systems that operate reliably in deployment do so by restricting the world they must handle, a pattern that recurs across Part IV. Driving systems are approved for a restricted operational design domain and add sensors and fallbacks to stay within it (Section 9.2); warehouse robots work in sites with fixed item sets and layouts (Section 9.5); surgical robots are mostly teleoperated and planetary rovers act within ground-supervised workflows (Sections 12.1 and 11.3); headsets track most reliably in previously mapped rooms, and the first home robots rely on remote operators for tasks they cannot yet perform (Sections 10.1 and 13.1). When the domain is widened, the world state is typically the first component to fail.
Guarantees end at the edge of the model. A constraint-based planner certifies collision avoidance only for the obstacles it knows about, and a driving system is validated only on anticipated scenarios (Sections 9.1 and 9.2). Learning moves these boundaries without notice: a model accurate on average can be confidently wrong in rare situations, its uncertainty estimates can lose validity when its use changes (Section 15.4), and an update that improves one behavior can silently degrade another (Section 9.2). No practical method yet states and verifies a property of a learned spatial system over a realistic domain. Legal responsibility is bounded the same way, since under conditional automation the manufacturer is responsible for driving only within a narrow domain (Section 9.2), and each extension reopens the question.
The large model cannot sit in the inner loop. A headset must respond to head motion within tens of milliseconds, a robot arm must update its state estimate at control rate, a drone must detect small objects in real time on onboard hardware, and a spacecraft must do similar work on radiation-hardened processors far weaker than ground computers (Sections 10.1, 9.1, 9.4, and 11.3). Large foundation models are much slower, even after work that accelerates them by predicting geometry internally and compressing spatial tokens [3020]. A practical architecture lets a small, fast component keep the world state current at sensor rate and lets the large model consult it when time permits, mirroring the fast–slow division that Section 7.3.5 described for tool-augmented reasoning. Whether that state is explicit, learned, or mixed is the question of Section 15.3; the deployment constraint is only that it be maintained at sensor rate and remain consultable.
A map starts going stale as soon as it is made. Furniture is moved, shop displays change, buildings are renovated, crops grow, and tissue shifts during surgery, so a deployed map must be updated continually, which Part IV found difficult in every domain. How a state should be revised is the algorithmic question of Section 15.3; the deployment problem is the cost and latency of keeping it revised and the synchronization of its copies across devices. Digital twins call the discrepancy between model and asset the liveness gap and treat closing it at scale as open (Section 11.1); lifelong SLAM must withstand months of appearance change (Section 6.3.5); shared augmented-reality maps are hard to keep current (Section 10.2); surgical navigation loses its preoperative registration once the brain shifts (Section 12.1). We suggest that deployed systems report how far their world state lags behind the world it describes, a quantity that is rarely measured.
Real data are scarcest where simulation is least trustworthy. Spacecraft perception is trained mostly on simulated images, and simulators do not reproduce the lighting of orbit, reflections from spacecraft materials, or radiation noise (Section 11.3); touch, force, and sound have the least data of all modalities and are also hard to simulate (Sections 7.3.8 and 9.1); clinical data are scarce because of privacy regulation and the cost of expert annotation (Section 12.1). In the first two cases simulation cannot compensate, since the simulator is least accurate for exactly the effects the missing data would capture. Simulations built from real measurements, data engines that target rare cases, and cross-institution training without sharing raw data all help, but none yet operates at the breadth that general spatial competence requires.
A persistent map is also an attack surface. Geospatial data raise issues of sovereignty and adversarial manipulation (Section 11.2), a compromised digital-twin server exposes the layout and vulnerabilities of the asset it mirrors (Section 11.1), and world-action models can be attacked through their visual inputs, with small perturbations producing large changes in behavior [3021]. These risks grow with the persistence of a map and the number of parties sharing it, turning the privacy concerns of Section 15.5 into security concerns.