How Autonomous Vehicles Deliver Packages

An autonomous delivery vehicle does not simply follow a line on a digital map. It observes the street, estimates what is happening around it, chooses a safe movement, and checks whether the choice remains sensible as traffic and conditions change. At the destination, it must solve a separate problem: stopping in a permitted, accessible place and getting the parcel into the customer’s hands.

Nuro’s R2 provides a concrete example of how engineering and regulation meet. In 2020, the U.S. National Highway Traffic Safety Administration (NHTSA) granted the company an exemption from certain Federal Motor Vehicle Safety Standards for the low-speed vehicle, which was designed to carry goods rather than people. The exemption was not a blanket approval for autonomous delivery everywhere. It addressed specified vehicle requirements; operation still depended on other applicable federal, state and local rules.

1. The delivery begins as a logistics assignment

Before a vehicle moves, a delivery system has to turn an order into an operational task. The order-management system identifies the parcel, its destination, any delivery instructions and a suitable vehicle or service. A warehouse or local depot prepares the load, records which compartment contains the package, and associates the parcel with the correct recipient.

Those steps sound routine, but errors here can become safety and service problems later. A package assigned to the wrong compartment may be impossible to retrieve at the curb. An address pin placed on the opposite side of a divided road could prompt an awkward or unsafe stop. A destination with restricted access may require the vehicle to use a designated loading area rather than pull up outside the front door.

Dispatch software therefore considers more than distance. It may account for vehicle capacity, battery or fuel state, operating boundaries, road restrictions, expected traffic, delivery deadlines and the time needed for each handoff. A vehicle may carry several parcels, but the order of stops matters: the shortest route is not always the route that best meets promised arrival windows or avoids difficult turns and stopping locations.

The assignment also needs a clear completion condition. “Arrived at the address” is not the same as “delivered.” A system may require confirmation that the vehicle has reached an approved stopping point, that the customer has authenticated successfully, and that the package compartment has been closed again before the job is marked complete.

2. Sensors turn the street into a working model

As the vehicle travels, it collects information about its surroundings. Depending on the platform, its equipment can include cameras, radar, lidar, ultrasonic sensors, satellite positioning and inertial sensors. These tools offer different strengths: cameras capture visual detail, radar can measure range and relative motion, and lidar provides depth measurements by sending out laser pulses. No single sensor answers every question in every condition.

Software combines measurements into an estimate of the nearby scene. This process is commonly called sensor fusion. The vehicle may track the position and movement of cars, cyclists and pedestrians, while also identifying lane boundaries, curbs, traffic lights, signs and temporary obstacles. It needs to estimate not only where an object is, but how confident it is in that estimate and how the object may move next.

Consider a parked delivery van partly blocking a lane. A camera might show its outline and brake lights; lidar can help estimate its shape and distance; radar may contribute information about motion and range. The system then has to decide whether there is enough room to proceed, whether it should wait for oncoming traffic, or whether the blockage requires a different route. A painted lane line alone cannot make that decision.

Perception is not perfect. Rain can obscure a camera lens, bright sun can reduce image clarity, roadworks can erase familiar lane markings, and a plastic bag can appear in sensor data as a more consequential obstacle than it is. Robust systems use multiple inputs and conservative decision rules, but they also need ways to recognize uncertainty. If the vehicle cannot confidently interpret a situation, slowing down or stopping may be safer than pressing ahead.

3. Maps guide the vehicle; live observations guide the next move

High-definition maps and route databases can provide useful context: road geometry, intersections, speed limits, access restrictions and known delivery zones. Satellite positioning and other localization methods help the vehicle estimate where it is. But a map is a reference, not a substitute for observing the actual street. A temporary road closure or newly placed barrier may not appear in the stored map.

The route planner chooses a path through the network, while a lower-level motion planner decides how to move along the next section of that path. It may calculate a safe trajectory and speed based on current traffic, road shape, nearby people and the vehicle’s own braking and turning limits. These decisions are repeatedly revised as new sensor measurements arrive.

That separation matters. The route planner may determine that the next destination is two blocks ahead, but the motion planner still has to manage the immediate approach to an intersection. If a pedestrian begins crossing, the vehicle must respond to the live event rather than insist on keeping its original schedule. A useful system is not one that never changes its plan; it is one that changes plans predictably and with an appropriate safety margin.

Autonomous vehicles also need operational boundaries. A service may be limited to selected roads, neighborhoods, vehicle speeds, weather conditions or hours of operation. These limits reduce exposure to situations the vehicle or operating organization is not prepared to handle. They can also make service easier to supervise and assess: a defined operating area offers a more tractable environment than an open-ended promise to travel anywhere.

4. Obstacle handling is a sequence of judgments

When something enters the vehicle’s path, the response depends on what the system believes it is seeing, how quickly the situation is changing and what safe options remain. A moving cyclist, a pedestrian at a crosswalk and a fixed construction barrier are all “obstacles” in a broad sense, but they call for different predictions and responses.

The vehicle’s software may weigh stopping distance, available road space, traffic behind it and the legal meaning of road markings. It may brake, pause, yield, or make a permitted path adjustment. In some cases, it will wait for the obstruction to clear. In others, it may need to stop and request assistance rather than attempt a maneuver it cannot confidently evaluate.

These choices are shaped by the operating design domain: the set of conditions under which a vehicle is intended to operate. A system designed for a mapped, low-speed route in a particular service area should not be assumed capable of handling every road, weather condition or temporary traffic arrangement. The operational limit is part of the safety design, not a footnote to it.

For fleet operators, an important question is what happens after a vehicle stops. A safe system needs a defined process for recognizing a fault, alerting the operations team, deciding whether the vehicle can continue, and arranging recovery if it cannot. Unplanned stops can affect traffic and customer service even when no collision occurs. Incident procedures should therefore cover both immediate hazards and the practical work of clearing or recovering the vehicle.

5. Remote operations support exceptions

Remote support is often misunderstood as a person driving every autonomous vehicle from a control room. In many operational models, onboard automation handles routine movement, while staff monitor vehicle status and respond to exceptions. A remote operator may review an alert, provide information about an unexpected road closure, or coordinate assistance. The precise role varies by company and deployment.

Remote assistance does not remove the need for the vehicle to behave safely while waiting. Networks can be delayed or unavailable, and a human reviewer may be handling several alerts. A vehicle should be able to reach or maintain a safe state if communications fail, rather than depend on an immediate remote command to avoid a hazard. Operators also need clear limits on what they may instruct the vehicle to do.

For logistics teams, remote operations have measurable implications. Staffing needs depend on how often vehicles request help, how long each case takes, and whether several events can occur at once. A service that appears economical when every trip is treated as fully autonomous may have different costs when actual road conditions, customer questions and access problems are included.

6. The curb is a planning problem, not an afterthought

Reaching an address does not guarantee a successful delivery. A vehicle may encounter a bus stop, a bike lane, a hydrant, a blocked loading bay, a driveway gate or a street where stopping is restricted. The curb is contested space, and the best point for a customer may not be a legal or safe place for a vehicle to wait.

Delivery planning can reduce this friction by using designated pickup points, verified access instructions and practical stopping locations. A customer may be asked to walk a short distance to meet the vehicle, particularly where a direct curbside approach is unavailable. That choice can improve safety, but it also affects convenience and accessibility. Operators should communicate the meeting point clearly and provide alternatives for customers who cannot reach it easily.

A simple example shows why the last few metres matter. Suppose a vehicle is assigned to an apartment building on a one-way street. The map pin is at the main entrance, but a loading zone is around the corner and the building’s front door is controlled by a call box. The delivery system must account for the legal stopping point, the customer’s ability to meet the vehicle, authentication, and a fallback if the recipient does not arrive. Navigation alone solves none of those steps.

7. The handoff needs security and a clear record

Once the vehicle is in an appropriate location, it has to release the correct parcel to the correct person. Delivery methods may include a one-time code, an app-based authorization, a compartment assigned to the order, or another identity check. The purpose is to reduce mistaken retrieval and protect goods while avoiding unnecessary collection of personal information.

The handoff process should make the vehicle’s status understandable. A customer needs to know where it has stopped, how to open the relevant compartment and what to do if the compartment will not open or the package is missing. The vehicle should not block a footpath or require a customer to step into moving traffic to retrieve an item. Clear interaction design is therefore a safety feature as well as a customer-service feature.

After retrieval, the system can record the event, close the compartment and return the vehicle to its route or next task. If the recipient does not appear, the operating rules may call for waiting for a defined period, notifying the customer, returning the parcel to a depot or attempting a later delivery. Those rules should be explicit; otherwise, a vehicle can occupy valuable curb space without a reliable path to resolution.

8. What the Nuro R2 exemption did—and did not do

Nuro’s R2 became a notable U.S. regulatory milestone when NHTSA granted an exemption in 2020 from certain Federal Motor Vehicle Safety Standards. The vehicle was designed to carry goods rather than occupants, so some standards written around human occupants did not fit its design in the same way. The exemption addressed specified requirements under federal motor vehicle safety law.

That distinction is central. A federal vehicle-safety exemption is not the same as permission to operate on every public road, a general endorsement of all autonomous technology, or a substitute for state and local approvals. Different authorities can have different responsibilities. NHTSA administers federal motor vehicle safety requirements; state agencies may regulate vehicle operation, testing or deployment; local governments may control matters such as curb use and traffic management.

For drone operators, FAA Part 107 is a familiar example of an operational rule set covering small unmanned aircraft. Road vehicles do not operate under Part 107; they face a different mix of motor-vehicle and traffic rules. The broader lesson is that “autonomous delivery approval” is not one universal permit. A deployment must identify which rules apply to the vehicle, the operator, the location and the particular activity.

The R2 milestone also illustrates why vehicle design affects regulation. A vehicle built without seats for passengers raises different questions from a passenger car, especially around occupant-protection standards. Yet carrying no people does not eliminate risks to pedestrians, cyclists and other road users. Speed, route, visibility, braking performance, operational oversight and incident response remain relevant to safe deployment.

9. Measure the service, not just the vehicle

A delivery program should be evaluated across the whole operation. Useful measures can include successful deliveries per route, on-time performance, energy use, vehicle availability, intervention frequency, failed handoffs, customer complaints and safety events. A high number of autonomous miles does not by itself show that the service is efficient or acceptable to the communities where it operates.

Operators should also distinguish between a vehicle’s technical performance and the performance of the service around it. A package can arrive late because of warehouse delays, an incorrect address, a blocked loading zone or a customer who misses a notification. Each cause calls for a different remedy. Treating every failure as a navigation issue can lead to the wrong engineering investment.

Data collection deserves careful boundaries. Cameras and other sensors may capture bystanders, homes and public spaces during ordinary operation. Organizations need policies for access, retention, security and disclosure, consistent with applicable law and the purpose for which information was collected. Protecting delivery records and vehicle-control systems from unauthorized access is also part of dependable service—not merely an IT concern.

What logistics teams should ask before a deployment

For a logistics professional assessing an autonomous delivery proposal, the most useful questions are specific to the operating area and the actual handoff:

• Where may the vehicles operate? Ask for the defined routes, road types, weather conditions, speeds and hours—not just a general coverage map.
• What happens when the system is uncertain? Understand its safe-stop behavior, remote-support process and recovery plan.
• How are stops selected? Check how curb restrictions, accessibility and customer meeting points are handled.
• Which approvals apply? Identify federal, state and local requirements for the vehicle, operator and service area.
• How is success measured? Request operational data on completed handoffs, exceptions, interventions and safety events, not only miles traveled.
• What is the fallback? A delivery service needs a plan for missed customers, blocked routes, vehicle faults and unavailable communications.

Autonomous package delivery is a chain of coordinated decisions, from assigning a parcel to choosing a safe stopping point and confirming retrieval. Sensors and planning software make the vehicle’s movement possible, but service quality depends just as much on bounded operating conditions, curbside procedures, regulatory compliance and exception handling. Nuro’s 2020 NHTSA exemption shows how one vehicle’s design can prompt a tailored regulatory decision; it also underscores why approval, safe operation and successful delivery must be treated as related—but distinct—questions.