Will Robots Finally Make Food Delivery Cheap?

12 Aug, 2026

Two pink Coco robots parked outside a Washington Thai restaurant carry a printed label that quietly contradicts the story usually told about them. The interesting question is not whether machines can deliver dinner, but who is still doing the work and where the savings land.

The Robot Outside Beau Thai Has a Driver

Sidewalk delivery robots arrived in the District this summer under a transportation permit that caps their speed and weight. The label printed on their flank says something the permit does not govern at all.

Two pink robots were parked outside Beau Thai, one of my favorite neighborhood restaurants in Washington, waiting with their orange safety flags up like a pair of oversized coolers that had learned to loiter. My first thought was not about robotics. I wondered whether they would make my Thai food cheaper, because anyone who uses DoorDash or Uber Eats knows how a reasonably priced menu turns into an unreasonable total once service fees, delivery charges, menu markups, taxes and a tip have been applied. Then I looked more closely at the panel on the side, below the phone number and the contact address that the District requires every such machine to display, and read the two words printed there: Human Operated.

Two pink Coco delivery robots parked on the sidewalk outside Beau Thai restaurant in Washington, D.C., each with an orange safety flag Two Coco personal delivery devices outside Beau Thai, Washington, D.C., August 2026. Elmer Yglesias (2026).

Human Operated

Coco's robots are remotely piloted. The company hires and trains employees to drive them from a dashboard, has said openly that it recruits people who are good at video games, and puts them behind controllers rather than steering wheels. Coco describes its pilots as trained staff rather than gig workers, and its chief executive has said they earn more than the United States minimum wage while declining to say how much more. The comparison worth holding in mind is Kiwibot, which drew criticism at Berkeley in 2019 when students learned its robots were being driven from Colombia at roughly $2 an hour, a rate the company noted was above the Colombian minimum. Remote operation makes the location of the labour a business decision rather than a physical constraint, and location is what determines the wage.

The saving, at least today, does not come from removing the human. It comes from changing what one human can be responsible for. A courier in a car is committed to a single order for the duration of a trip, plus the drive to the restaurant, the search for parking and the empty return leg. A pilot at a dashboard has none of that overhead and can switch between orders, neighbourhoods and cities between deliveries. Automation of that kind improves utilisation rather than eliminating labour, and the cost curve depends entirely on how many robots one person can eventually supervise instead of drive.

Coco has been explicit that supervision is the destination. When it launched Coco 2 in February 2026, the company framed the new machine as the shift from human-guided robotics toward full autonomy, powered by what it calls the largest dataset of sidewalk robot operations, accumulated across roughly a million miles of flooded Miami streets, Chicago snow and Los Angeles traffic. The pilots are, in the ordinary way of these arrangements, generating the training data for the system that will need fewer of them.

DoorDash provides the same lesson from the opposite direction. Its purpose-built robot, Dot, navigates roads and bike lanes on its own at up to 20 miles per hour, and in July the company was reported to be offering Dashers about $5 to drive to a restaurant, collect an order and place it inside a waiting Dot. The vehicle that needs no driver still needs someone to open the lid. A robot on a sidewalk is a proxy for autonomy rather than evidence of it, and the difference matters for anyone trying to forecast what the delivery will eventually cost.

Where the Dollar Goes

The most quoted number in this debate comes from a Barclays note reported by Reuters on 15 April 2026. Autonomous delivery in early-adoption markets with high labour costs currently runs about $5 to $7 per drop, which is already $3 to $4 cheaper than a rider, and Barclays argues the figure could fall to roughly $1 in the long run, implying savings of $8 to $9 against rider delivery in expensive labour markets. Penetration today is under 1% of global food delivery orders, and the bank projects around 10% by 2035.

The sentence that matters most in that note is the one about who collects. Assuming roughly $4 of saving per drop at long-term penetration, Barclays estimates the shift unlocks about $16 billion in annual global profitability for delivery platforms. Nothing in the framing is improper; sell-side research answers the question its readers are paying to have answered, and those readers own DoorDash, Uber, Meituan and Prosus rather than dinner. It is simply the wrong document to consult about the customer's bill, and the unit of account it uses tells you where the industry currently expects the money to settle.

The early evidence points the same way. DoorDash states plainly on its own Dot page that delivery fees for a robot delivery are the same as they are with a Dasher. Cost falls, price holds, margin absorbs the difference. Nor does the courier's wage dominate what a restaurant pays: DoorDash's partnership plans charge merchants 15%, 25% or 30% of the subtotal for marketplace delivery, and a free courier does not touch a commission calculated on the food. The restaurants that tolerate these platforms are paying for demand and for the interface, not primarily for the trip.

What would move consumer prices is competition rather than generosity, and competition requires enough autonomous capacity that one platform can undercut another visibly. A neighbourhood in which one service quotes $7 to carry a $20 meal six blocks while another quotes $2 for the same food in the same time is a neighbourhood where the first price cannot hold. Below a certain fee the composition of orders changes too, because a $14 lunch that currently arrives as a $25 total is an order that mostly does not get placed, and cheap short trips create demand that does not exist today rather than merely repricing the demand that does.

The District as a Test Bed

Washington opened this door in stages. The District Department of Transportation announced in July 2026 that it had issued personal delivery device permits to Serve Robotics and Coco Robotics, joining Robot.com, formerly Kiwibot, as permitted operators in the core of the city; both new entrants deliver for Uber Eats, and Coco also works with DoorDash and directly with merchants. Coco announced its Washington expansion on 15 July and says it is scaling toward thousands of vehicles across its markets by the end of 2026. The permit terms are specific about the machine: no more than 10 miles per hour, no more than 275 pounds excluding cargo, operation confined to sidewalks, crosswalks and alleyways, pedestrian traffic rules with a standing obligation to yield, a documented process for malfunction and removal, and a field test of interactions with pedestrians with disabilities before approval.

Density is what makes the experiment sensible here. Coco's own argument, that sending a 3,000-pound petrol car to carry a sandwich down the block is an absurd use of a vehicle, is hard to dispute in a neighbourhood where a restaurant such as Beau Thai can reach most of its customers within half a mile. Direct merchant relationships also make it plausible that a restaurant eventually keeps two or three machines outside during the dinner rush the way it once kept its own driver, which would blur the line between takeout and delivery rather than replacing one with the other.

The dispatch layer is where the platforms are actually competing. DoorDash's Autonomous Delivery Platform is described by the company as an AI dispatcher matching each order to a Dasher, a Dot, a sidewalk robot or a drone according to speed, cost and location. Coco does not need to displace DoorDash or Uber Eats under that model; it needs to become a selectable mode inside them, which is a considerably easier business to build. The durable asset of these companies was never the courier network. It is demand, routing, merchant relationships and the checkout screen, and a mode swap underneath leaves all of that intact.

Governance runs a step behind. The permit governs the device thoroughly and says nothing about the person steering it, which is the one variable that determines both the current cost of a delivery and the working conditions attached to it. A regulator reading the terms and conditions can tell you the maximum weight of the machine on the sidewalk outside Beau Thai. Nothing in the file establishes where its driver is sitting, how many machines that driver is responsible for, or what the hour is worth.

The Distance to a Dollar

Three things would have to be true before a delivery costs $1. Supervision ratios would have to rise sharply, so that one person oversees a fleet rather than drives a unit. The handoff problem would have to be solved at both ends, because a robot that requires a paid human trip to load it has moved the labour rather than removed it. And the fleet would have to survive the ordinary attrition of cities, which means weather, kerbs, theft, vandalism, battery replacement, insurance, remote monitoring and the cost of retrieving a machine that has stopped in the wrong place.

The historical analogies usually offered here are more ambiguous than they sound. Cash machines did not eliminate bank staff, self-checkout did not empty grocery stores, and lifts did not remove building employees, but in each case the work was redistributed and repriced rather than simply added to elsewhere, and the people who had been doing the automated part were rarely the people who got the new roles. Simple, short, single-bag deliveries may well be a large share of urban restaurant orders, and a system that takes those trips out of circulation leaves human couriers with the difficult remainder: stairs, locked lobbies, heavy orders, bad weather and vague instructions. Both consequences arrive together, and a customer saving $4 on dinner and a courier losing the easiest run of the evening are the same transaction viewed from two ends.

The scene outside Beau Thai looked futuristic enough that I stopped to photograph it. The more interesting artefact turned out to be the printed panel rather than the machine, because it was the only part of the assembly that told the truth about the labour. Someday the app may simply say $1.99 and make the transportation decision invisibly, which is what success looks like for a technology, and by then nobody will read the side of the robot. Until then the useful question is not whether the machine can find my building. It is who is on the other end of the controller, and what happens to the difference between what they cost and what I pay.

My blog