Short answer
Last-mile handling data covers carrying parcels from vehicle to door, climbing steps, ringing or knocking, and placing items where recipients expect them. It is recorded in public and residential spaces, so bystanders, homes and worker monitoring raise permission questions that staged routes and depot recordings reduce but do not remove. Video supplies no robot actions or load forces.
The buyer's problem
Delivery sits at the boundary between a logistics operation and the public. Inside the depot the environment is controlled. On the street, the task is highly variable: a van parked at a bad angle, a gate with a latch, steps with no rail, a dog, an unclear house number, a package that will not fit a letterbox, rain. The worker makes dozens of small decisions about where a parcel may be left safely.
A robotics buyer needs examples of those decisions, and the footage that shows them most naturally was often recorded in front of people who did not agree to be recorded. That tension decides which supply route makes sense.
Task anatomy
- Load-out. Scan and sequence parcels in the vehicle; pick the next stop's items; arrange by size and fragility.
- Vehicle to door. Lift from shelf or cargo area, carry one or several parcels, use a trolley or hand truck, handle the vehicle door and a gate.
- Approach. Find the entrance, climb steps or ramps, avoid hazards, locate a doorbell or house number.
- Hand-over or drop-off. Knock or ring, wait, hand to recipient or choose a safe place; confirm with photo or signature.
- Exceptions. No answer, locked entry, oversized parcel, damaged package, redirect to neighbor or locker.
- Return. Re-enter the vehicle, close the cargo area, move to the next stop.
For carrying techniques and load reaction, see manual load handling and carrying. For steps and ramps specifically, read workplace movement: stairs and ladders.
Why privacy dominates this task
Three groups appear in outdoor and doorstep footage:
- Bystanders and residents. Faces, homes, vehicles and conversations. The EDPB guidelines on video devices discuss the conditions under which personal data from video devices may be processed and the position of people who happen to be in view (EDPB, 2020). Outside the US, the legal position for imported footage varies.
- The delivery worker. Body-worn or vehicle cameras used to monitor drivers are a workplace monitoring issue. The UK ICO guidance on monitoring workers recommends a data protection impact assessment for monitoring likely to result in a high risk to workers' rights and freedoms (ICO, 2023). Footage collected for safety or dispute resolution was not necessarily collected for training.
- Recipients. A porch scene shows a home, address, packages and perhaps a person's schedule. Ask what notice or consent, if any, applied to recipients.
This is not legal advice. The questions are for the supplier, its counsel and your own. Blurring faces and plates does not by itself settle them, because houses, voices and routes can still identify people. See redaction in demonstration video for what redaction costs the training signal.
Which route fits each part
| Part of the task | Realistic source | Reason |
|---|---|---|
| Depot load-out and sorting | Existing depot video, or staged recording | Controlled space with employees; consent is easier to document |
| Vehicle loading and unloading | Staged or depot-yard recording | Few bystanders |
| Carrying and stairs | Staged route on private or leased property | Controlled steps and surfaces |
| Doorstep hand-over | Staged with consenting "residents" | Limits exposure of real residents' data |
| Exceptions and public sidewalks | Hardest to source; staged approximations | Consent from public bystanders is rarely practical to obtain; treat as a diligence question |
Delivery operators sit alongside the 3PLs that handle first- and middle-mile work; see 3PL providers as data suppliers for the depot side.
Sourcing
Existing footage. Delivery operators may hold depot cameras, dashcams and driver-facing cameras. Depot footage is the cleanest candidate, because it shows employees in a workplace and not the public. In-vehicle and body-worn footage shows the street; its original purposes (safety, claims, training the driver) are narrow, and reuse needs a clear basis from the operator.
Commissioned recording. Staged sessions use a leased building, a mock porch, a campus or a closed lot; participants playing recipients should be told what is recorded and agree to it; ask for the consent records. You choose parcel sizes and weights, step geometry, door types and exceptions. Real-street capture should be limited to what the operator and local rules allow, and treated as a special case under the recording protocol for live work sites.
If you can say which parts of the route matter, describe the last-mile tasks and permitted uses to SourceX and it will investigate whether US businesses hold suitable material and discuss whether a staged collection is feasible.
Illustrative package
# Illustrative only - synthetic values, not an available dataset
package: last_mile_staged_pilot
sites: depot_yard_040 + leased_house_frontage_041 (mock porch, 5 steps, gate)
couriers: 6 (courier_K01..K06), written consent, DPIA-style review documented
participants_as_recipients: 8 consenting adults, ids resident_01..08
parcels: 5 size classes, weights logged per parcel, 2 fragile-labeled
routes: [van_to_porch_carry, stairs_with_parcel, ring_wait_handover, no_answer_dropoff, oversize_to_locker]
views: head_mounted; body_chest; fixed_porch; fixed_vehicle_rear
labels: step, parcel_id, carry_mode (hands/trolley/two-hand), outcome (handed/dropped/left_safe/returned)
weather_note: dry / rain / low_light
excluded: forces, robot actions, real residents, public sidewalks
rights: training and evaluation; no redistribution
Quality and acceptance pointers
- Parcel diversity. Count sizes, weights, shapes and wrapping. All-cardboard-boxes is a narrow set.
- Exception coverage. No-answer, locked gate, wrong address and oversize cases should be present, not only success.
- Route geometry recorded. Step count, rise, rail presence and door type belong in metadata.
- Weather and light. Rain, glare and dusk affect perception; record them per session.
- Evaluation split. Hold out a property type, a courier and a parcel set. See mobile manipulation demonstrations for the combined navigation-and-handling framing.
What this data cannot support
- Navigation policy. Human walking video is not a robot's navigation data. Obstacle avoidance, path planning and curb handling need robot-side sensing.
- Robot actions and loads. There are no joint commands, no payload limits, and no measured parcel weight unless logged.
- Real public behavior. Staged bystanders behave differently from the public. Interactions with strangers, pets, traffic and weather are thin.
- Safety and liability rules. What a robot may leave at a door or hand to a minor is a policy matter outside any dataset.
- Address and map knowledge. Finding a unit in a complex needs maps and local knowledge, not demonstration clips.
Questions to ask suppliers
- Which footage comes from depots, vehicles, body-worn cameras or staged sessions?
- What was each camera originally installed for, and what notice did drivers and the public receive?
- Were recipients, bystanders or homes ever in frame, and how was that handled?
- Has a monitoring-impact assessment been done for any worker-facing cameras?
- Are parcel weights and sizes logged?
- Are exceptions, such as no-answer or dropped parcels, kept and labeled?
- What route geometry and weather conditions are covered?
- Is audio included, and does it contain conversation?
- What did couriers consent to, including use of movement data in model training?
- Can a property type or depot be held out for evaluation?
Discuss your requirements
Discuss your humanoid training-data requirements with SourceX by describing the delivery tasks, environments and permitted uses. SourceX reviews the requirement and follows up about potential sources; a request does not guarantee a match, and nothing is contracted until a supplier agrees. See the workplace task demonstration hub, the humanoid training data guide, and the catalog page on supply chain and logistics data.
Describe the workplace demonstration data your robotics project needs
Sources
- Guidelines 3/2019 on processing of personal data through video devices (Version 2.0) — personal data in video recordings of public and residential spaces.
- Employment practices and data protection: monitoring workers — monitoring of workers and DPIA recommendation.