Notes from the edge

From camera to agent: forklift safety in 30 minutes

Every warehouse we visit has the same poster on the wall: "STAY CLEAR OF FORKLIFTS." And every warehouse has near-misses anyway, because posters don't watch aisles. Here's the full build of a system that does — using one Spatial Robot LiDAR, the forklift-safety model, and a ten-line agent.

1. Mount and enroll

The Spatial Robot's SPAD dToF covers 8 m of choke point — the intersection where near-misses actually happen — from a single pole mount, unfazed by skylight glare. One cable, then visionlibra fleet enroll cam-047 and it appears in the dashboard.

2. Load the model

The forklift-safety model classifies forklifts and people from pure depth, tracks both, and emits a person_near_forklift event when trajectories converge. It runs on-device in real time.

3. Write the agent

Subscribe to the event; wire three actions: Slack the safety channel, log a structured record to Postgres, and generate an incident report with the depth clip attached. That's the whole program — it's the exact code on our Developer page.

4. Watch it work

Events land in Slack in under a second. The dashboard heatmap shows convergence hotspots by aisle, which — in every deployment so far — moved a pallet staging area within the first week. That's the difference between AI and an AI agent: the first one sees, the second one changes the building.

Staged OTA for 100 cameras without a single bricked device

Firmware updates are where camera fleets go to die. Our rollout engine updates in waves — 10%, then 50%, then all — with automatic health verification between stages. A device that fails post-update checks rolls back to the previous image on its own; the stage halts and the dashboard flags the diff.

The A/B partition scheme means there is no unbootable state, and the fleet CLI makes the whole process one command: visionlibra ota rollout v2.4.2 --stage 10%. The result across our pilot fleets: 3,000+ device-updates, zero truck rolls.