What to Ask a Roadside Compute Vendor When the Map Might Be Lying
Ask any vendor of roadside compute and dynamic map services what the vehicle does when a roadside unit sends wrong data with a valid signature. A unit that goes fully offline is easy to handle. The harder case is a unit that keeps transmitting after it has lost its sensor feed or its backhaul, and that case decides whether the architecture fails safely or misleads the car. If you have a shortlist and little bench time, treat each question below as a test for one specific risk, and judge the answer by how precise it is rather than by what the datasheet claims.
Who declares the data dead?
Ask where staleness is detected: at the RSU, in the map aggregation backend, or in the vehicle. The risk behind the question is a failure seen across deployments. The RSU loses its camera or radar feed, or the link to the map service, yet its message scheduler keeps sending the last known object list with fresh timestamps, because the timestamp is applied when the message is sent rather than when the scene was sensed. The security credentials are still valid, so the vehicle's message verification passes, and the stale hazard, or the hazard that is missing from the list, enters the fused world model. A credible vendor stamps each object with its sensing time, reports a health flag for each sensor, and lets the vehicle apply its own age limit, for example discarding any dynamic layer object whose sensing timestamp is more than 300 ms old relative to the vehicle's synchronised clock. If the vendor cannot say which clock domain that timestamp belongs to, or how drift is bounded when GNSS timing is lost, the age limit means nothing.
The floor must not depend on the roadside
Reject any architecture in which the vehicle's minimal risk manoeuvre depends on a message from the roadside, the map backend or the cellular link: infrastructure can extend the operational design domain, but it should never be what the vehicle needs in order to stop safely.
Redundancy that shares a fate
Vendors will show dual RSUs per intersection, redundant edge servers and active-active map backends. Ask what those redundant paths have in common: the power cabinet, the fibre trench, the time source, the certificate authority, the software build. The real risk is common-cause failure, and an over-the-air update that pushes a parser bug to every RSU in a corridor at once defeats any amount of hardware duplication. Ask for the staged rollout policy and the rollback trigger, and ask whether the vehicle can pin a map schema version or reject one it has not validated. Functional safety standards such as ISO 26262 and ISO 21448 give you the vocabulary for this conversation, but neither one certifies a roadside cabinet, so ask for the vendor's own dependent failure analysis instead of accepting a compliance logo.
Spending scarce lab hours
With limited lab access, choose the faults that tell you the most. Pull the sensor feed while the radio stays on, cut the backhaul while local perception keeps running, and skew the RSU time source. For each fault, check whether the vehicle stack lowers its confidence, widens its margins and falls back to onboard perception and its static HD map, or whether it keeps planning on stale data. A vendor that passes all three on the bench has dealt with the failures most likely to cause field incidents. The cost is availability: tight age limits and strict health checks will discard roadside data more often where coverage is poor, so you still have to decide how much of the promised sensing range you will give up to keep the system safe when a part fails.