Edge AI needs a decision plane and a coordination plane

The architecture behind moving drone vision decisions to the edge while preserving fleet-wide control, observability, and model lifecycle.

Edge intelligence succeeds when local autonomy and central coordination have separate responsibilities and a deliberate failure contract.

Cloud-only vision breaks at the mission boundary

Sending every frame to the cloud can work in a laboratory with stable bandwidth. A field system experiences variable links, shared networks, moving devices, thermal limits, and moments when a delayed detection is no longer useful. The architecture has to remain valuable when the connection is degraded rather than treating that condition as exceptional.

The response was not to move the whole platform to the drone. It was to divide responsibility by time sensitivity and scope.

System design: two planes, two failure contracts

The edge decision plane

The decision plane receives frames, runs optimized inference, applies mission policy, and emits compact events. It knows enough to decide what deserves attention now.

• Owns: local inference, frame sampling, confidence policy, thermal behavior, evidence selection
• Does not own: fleet history, long-term analytics, global scheduling, the full model lifecycle

Optimization includes more than model quantization. Frame sampling, resolution, region-of-interest processing, confidence policy, thermal behavior, and evidence selection all affect whether the edge system is operationally useful.

The cloud coordination plane

The coordination plane manages missions, device identity, configuration, model rollout, cross-device correlation, operators, and long-term analysis. It receives events by default and requests or accepts richer media when required.

• Owns: mission state, device identity, model rollout, cross-fleet correlation, long-term analysis
• Does not own: the millisecond-level decision of whether this frame deserves attention right now

This division reduces bandwidth without making the fleet invisible. It also gives each plane a cleaner failure contract, shown below for a connectivity drop.

Operate models like fleet software

A model deployed to an edge fleet needs version inventory, staged rollout, device compatibility checks, performance telemetry, rollback, and drift investigation. Accuracy measured on a workstation is insufficient if the model overheats a device, misses its latency budget, or behaves differently across camera conditions.

• Canary model versions on a small device ring
• Correlate model version with latency and event distribution
• Preserve a last-known-good rollback
• Test offline and reconnect behavior as product scenarios

From production: the numbers that justified the split

The 80% bandwidth reduction measured on CGC Sentinel didn't come from a single clever compression trick. It came from refusing to send the full video stream by default and treating it as evidence retrieved on demand instead of a permanent broadcast — the "events before raw video" decision, with the accepted trade-off that less raw footage is centrally available in exchange for a major bandwidth and response-time gain.

The roughly 60–70% reduction in decision latency came from removing the round trip entirely for the decision that actually mattered: whether this frame deserves attention right now. That number only holds because the edge decision plane was never asked to also own fleet history or global scheduling — the moment a local decision has to wait on a cross-device correlation, the latency budget that makes edge inference worth deploying disappears.

Model rollout used a canary ring on a small set of devices before a wider release, with a last-known-good rollback kept ready at all times — because a model regression on a moving drone with a thermal budget and a live mission is a materially different risk than a regression behind a web feature flag. The cost of getting the rollback path wrong isn't a bad session; it's a device that keeps making a bad call in the field until someone notices.

What to carry forward

• Place decisions according to time sensitivity
• Transmit events and selective evidence by default
• Define offline behavior explicitly
• Manage edge models as a fleet lifecycle
• Keep the rollback path as tested as the rollout path

Applied in the portfolio

• CGC Sentinel — Edge computer vision for drone operations, moving decisions closer to the camera and reducing network pressure.