Problem
Last-mile logistics is the most expensive segment of the delivery chain. Industry estimates put it at 40–50% of total delivery cost for parcel shipping, despite covering only the final 5–10% of the distance. Inside the building — warehouse aisles, fulfillment-center mezzanines, multi-floor distribution hubs — that cost gets even worse, because the final 100 meters often requires a human picker walking the parcel from a pallet to a dispatch bin.
Conveyor systems and AMRs (autonomous mobile robots) solve some of this, but they share two weaknesses: they’re confined to the floor (so they can’t go over shelves or through narrow vertical corridors), and they require a structured environment (markers, charging stations, predictable pallet layouts). Retrofitting an existing warehouse to support them is a six-figure capital project before the first robot rolls.
Agitation
The pieces of this problem that classical approaches haven’t solved:
- Vertical traversal. Warehouse picking is genuinely 3D — pallets stack 4-6 high. An AMR can’t reach the third row without a human or a lifting mechanism. Conveyors can but require fixed routes.
- Unstructured environments. Real warehouses change weekly: pallet positions shift, racks get rearranged for seasonal layouts, temporary boxes block aisles. Hard-coded path planning breaks; humans don’t.
- GPS denial. Indoor positioning systems exist (UWB beacons, vision markers) but they’re an infrastructure investment per facility. A drone that doesn’t need infrastructure is genuinely cheaper to deploy.
- Cost floor. Industrial mobile robots are $30K–$100K per unit. For a single-parcel handoff worth $2–5 of labor savings each, payback periods are years.
Solution
A small indoor drone (~600 g all-up, 30 cm motor-to-motor) that flies vertically over the racks, picks up a parcel from a designated handoff zone, and delivers it to a dispatch bin — without any beacons, markers, or environmental modifications. The cost target is under $1,500 per unit at quantity (the BOM puts us in that range), making the payback economics work even at modest utilization.
Three technical choices make this feasible:
- Layered low-cost sensors (TF-Luna + VL53L0X array + PMW3901 optical flow) covers obstacle avoidance, altitude hold, and lateral drift correction without a 2D scanning lidar. Total sensor cost: ~$80.
- RL policy trained in warehouse_v2 handles the cluttered-environment piece. Classical planners do fine when aisles are predictable; the policy keeps working when they aren’t.
- No infrastructure dependency means deploying a fleet is just charging stations and a network — no marker survey, no UWB rollout.
What we’re not claiming
Outdoor delivery (drone-to-doorstep across kilometers) is a different problem with different constraints (regulation, range, payload). We’re not solving it. Multi-kilogram payloads also aren’t the target — this is small-parcel logistics, the kind of work that currently requires a human walking it the last 50 meters. And the policy isn’t magic: out-of-distribution layouts (e.g., a totally novel warehouse it’s never seen during training) will need a domain-randomization pass before deployment.
Where to go next
- 3-rangefinder sensor stack — why the cost structure works
- RL vs lawnmower benchmark — concrete numbers on indoor coverage performance
- Monitoring & inspection — adjacent use case with overlapping hardware