May 2026 — Where we are

Already concrete

• Costed procurement tracker; ~80% of Phase 1 P0 parts ordered
• Browser viewer with two georeferenced park sites
• Simulation + Gazebo SIL setup
• First sustained perception benchmark on bare Pi 5 1 GB
• Site-class business model in the operational docs

Still to prove

• Physical pickup on a real piece of trash
• Bounded route demo with before/after proof
• Field-condition thermal envelope (summer enclosure)
• Operator handoff UX for low-confidence detections
• Repeatable economics on a real recurring corridor

What worked this month

• First sustained perception benchmark on a bare Pi 5 1 GB — 1 hour, 7,649 inferences, 0 throttle, 0.05 % drift, 73.6 °C peak. Latency budget for safety planning: ~220 ms per inference.
• Procurement closed on a usable parts list for Phase 1 build.

What didn’t

• Bench Pi 5 originally ran undervolted off a laptop USB-C cable; benchmarks dropped Wi-Fi during inference. Fixed with a proper 27 W PSU.
• Original front caster SKU returned a 404 at order time; rerouted to Service Caster direct.

Next 30 days

• First chassis assembly
• Add accuracy (mAP) measurement to the benchmark harness
• Stand up the Trash Trail API + minimal PWA