Delivery robot

A delivery robot is a wheeled, uncrewed ground vehicle that carries goods over the last mile of a delivery, from a store, restaurant or depot to a customer. The category is not one thing. It splits into machines that behave like pedestrians and machines that behave like vehicles, and the split governs what they weigh, how fast they go, which surfaces they use, which agency regulates them, and what happens when one hits somebody.

Sidewalk robots are the familiar cooler-sized machines: footpaths, walking pace, typically 3 to 6 mph, and in the United States most operate under state personal delivery device (PDD) statutes that classify them as something other than a vehicle. Road-going delivery robots are larger, run at 15 to 25 mph, and share bike lanes and neighborhood streets with traffic. A third class, low-speed autonomous delivery vehicles, are road-only machines built and regulated like cars without occupants. This article covers all three. Warehouse machines are covered at autonomous mobile robot, aerial last-mile delivery at drone delivery.

The taxonomy is routinely garbled in press coverage. DoorDash's Dot is a 350-pound, 20 mph road machine that DoorDash's own engineers explicitly contrast with sidewalk robots ("Traveling up to 20 mph, Dot can ship items faster and much further than sidewalk robots"), and whose dispatch layer treats Coco Robotics' sidewalk robots as a separate mode entirely ^[1][2]. Newsweek, DroneDJ and TechRepublic called it a sidewalk robot anyway, each in passing and each ten months after launch, during the July 2026 coverage of DoorDash's drone programme ^[44][45][46]. TechCrunch did both: its launch coverage drew the distinction correctly, noting that sidewalk robots "can't drive on roads or at high speeds like Dot can" ^[27], while its July 2026 piece called Dot "the autonomous sidewalk delivery bot" ^[47]. Nuro's R2 is neither: a 25 mph road vehicle that received a federal motor-vehicle safety exemption ^[3].

Scroll sideways for more →

History

2014 to 2018: the first wave

Starship Technologies was founded in Tallinn on 3 July 2014 by Ahti Heinla and Janus Friis, with a first prototype built the following month; Heinla, still chief executive and chief technology officer, has said the idea came out of a NASA competition to build a Mars sample-collection robot ^[4]. Starship's own material is inconsistent about their Skype roles: the company profile calls them "Skype co-founders," while its June 2026 press release is more precise, "Ahti Heinla (chief architect of Skype) and Janus Friis (co-founder of Skype)" ^[4][48]. Starship stealth-launched its first robot, the 6C, in 2015; by 2018 it had a commercial service in Milton Keynes, UK, and claimed 22,500 cumulative deliveries ^[4].

The San Francisco startup Marble, co-founded by Matt Delaney, Jason Calaiaro and Kevin Peterson, built robots carrying lidar, cameras and ultrasonic sensors driven by NVIDIA Jetson TX1 modules, navigating against high-resolution 3D maps. In April 2017 Marble began a food-delivery pilot in San Francisco with Yelp Eat24, and it sent a human chaperone alongside every robot ^[5]. Kiwibot, founded by Felipe Chavez and backed early by the UC Berkeley SkyDeck Fund, took the cheaper route with small campus machines ^[16].

Amazon entered on 23 January 2019 with Scout, a six-robot pilot in Snohomish County, Washington. Its own announcement was candid about the autonomy: "The devices will autonomously follow their delivery route but will initially be accompanied by an Amazon employee" ^[6]. Starship launched its first US campus service the same month, at George Mason University ^[4].

2020 to 2022: the pandemic and the shakeout

COVID-19 made contactless delivery briefly indispensable and the numbers moved sharply: Starship reports 957,000 cumulative deliveries by the end of 2020 and 2.4 million by the end of 2021 ^[4]. In China, JD rushed autonomous delivery vehicles into quarantined neighborhoods in Wuhan and Meituan deployed them into sealed-off Beijing communities ^[7].

The same period produced the sector's first graveyard. Caterpillar announced in June 2020 that it had acquired select assets and hired employees from Marble Robot, redirecting the team toward construction, quarry, industrial and waste autonomy; Peterson joined Caterpillar as a technologist ^[8]. Amazon ended Scout field tests in October 2022, closing a programme that had grown to roughly 400 people and expanded from Washington to Southern California, Atlanta and Franklin, Tennessee; a spokesperson said Amazon had "learned through feedback that there were aspects of the program that weren't meeting customers' needs" ^[9].

Two survivors have complicated lineages. Serve Robotics began in 2017 as the robotics division of Postmates, later Postmates X, and showed its first robot in December 2018 ^[10][11]. Uber acquired Postmates in 2020 and spun the unit out on 2 March 2021, with Ali Kashani as co-founder and chief executive and Uber retaining an investment ^[10]. Nuro was founded in 2016 by Jiajun Zhu and Dave Ferguson, both former principal engineers on Google's self-driving programme, now Waymo ^[12].

2023 to 2026: consolidation, pivots and a road-going turn

Nuro's arc is the clearest cautionary case. On 6 February 2020 NHTSA granted it the first exemption ever issued for a driverless vehicle, allowing up to 5,000 R2 units over two years to operate without mirrors or a windshield ^[3]. After layoff rounds in 2022 and 2023, Nuro announced on 11 September 2024 that it would stop owning and operating a delivery fleet and instead license its autonomy stack, having logged over a million autonomous miles across Arizona, Texas and California through partnerships with Uber Eats, Domino's and FedEx ^[13]. Its company page now describes it as supplying "fully autonomous driving technology for mobility and robotaxi services" ^[12].

Cartken, founded in 2019 by former Google engineers from the Bookbot project, also moved away from consumer sidewalk delivery, and now markets to manufacturing sites, warehouses and pharmaceutical campuses rather than restaurants ^[14][15]. Kiwibot rebranded as Robot.com on 31 October 2025, claiming more than 1.7 million completed tasks and over 500 robots deployed globally, largely through a campus food-service partnership with Sodexo running since 2021 ^[16]. Serve acquired Diligent Robotics in early 2026, adding hospital service robots to sidewalk delivery ^[17].

The largest single retreat came from the market leader. On 4 June 2026 Starship announced it was winding down its US university campus operations, the business that had made its name in the United States after the 2019 George Mason launch, and redeploying "over 1,200 robots from the U.S. campuses' fleet" to grocery retailers in Europe and the United States. Heinla framed it as focus rather than failure: "it's time for us to focus on the vertical we feel will have the most value." Starship said service to campus partners would continue through the 2026-2027 back-to-school season ^[48]. The company that pioneered the American campus delivery robot has left the category it created.

DoorDash's Dot, unveiled on 30 September 2025 and built in-house at DoorDash Labs, is the sector's clearest bet on the road-going class ^[1][2].

How delivery robots work

Perception

Two design philosophies compete, and the split is driven mostly by cost. Camera-primary stacks were the underdog position in 2019, when lidar units cost thousands of dollars each. Refraction AI co-founder Matthew Johnson-Roberson put the economics bluntly when the REV-1 launched: "It doesn't make sense economically speaking to use a $10,000 lidar to deliver $10 of food." The REV-1 used twelve cameras plus radar and ultrasound, and no lidar ^[18]. Cartken markets the same choice today, a "Camera-First" system giving "precise, scalable automation without costly LiDAR" from depth and colour cameras with sensor fusion ^[14]. Starship's robots carry twelve cameras, radar, ultrasonic and time-of-flight sensing ^[20]; its FAQ describes obstacle detection "using a 'situational awareness bubble' around it," in which the robot "reduces its speed" for an object adjacent to it and comes "to a complete stop" for one directly in front ^[19]. Avride goes the other way, layering lidar that detects objects up to 60 metres out with colour cameras and ultrasonics ^[21].

The interesting movement is in the middle. Dot runs a vision-primary stack of 8 external cameras giving 360-degree coverage, one interior camera to verify delivery quality, and 4 inexpensive radar units. It also carries three high-resolution lidars, but DoorDash says plainly that these "are being replaced by inexpensive automotive-grade lidars resulting in a low-cost sensor stack to enable rapid commercial scale" ^[1]. That is the industry trajectory stated out loud: not lidar versus no lidar, but expensive lidar versus commodity lidar, so the sensor bill can be amortised across a $12 burrito rather than a $60,000 car.

Localization and mapping

Sidewalks are a harder localization problem than roads, for a reason easy to miss: they have almost no machine-readable structure. There are no lane markings, no consistent width, no standard curb geometry, and the objects that define the corridor (parked cars, wheelie bins, A-boards, scaffolding) change hourly. GNSS position is simultaneously degraded by the buildings and tree canopy that line those same corridors, so satellite fixes adequate on a highway drift by metres in an urban canyon.

The universal answer is a prior map. Marble navigated against high-resolution 3D maps of sidewalks and buildings in 2017 ^[5]; Starship combines proprietary mapping with computer vision and GPS and maps crossing points in advance, so that "robots will only cross roads at agreed crossing points" ^[19][20]. Cartken's system reportedly does not depend on GPS at all, which is what lets the same stack run indoors ^[15]. In practice these are variants of map-relative SLAM: visual and lidar features are matched against the stored map to recover pose, with wheel odometry and an inertial measurement unit filling gaps between fixes. The consequence is that a service area is a hand-built asset. Expanding to a new neighbourhood is not a software update; it is a survey.

Planning and behaviour prediction

DoorDash has published the most explicit account of a delivery planning stack in the sector: "Our stack combines deep learning and search-based algorithms to find a safe, smooth path through a complex world. Deep learning allows us to understand how other road users act, and, in turn, how we should drive in complex situations. The search acts as a safety net and ensures the robot navigates the environment safely, smoothly, and predictably" ^[1].

The reason for that two-part architecture is not performance but assurance. A learned prediction model is good at the thing classical motion planning is bad at: guessing what a cyclist, a van or a distracted pedestrian is about to do. But a learned policy carries no guarantee that under some unseen input it will not propose a trajectory through a person. A search-based planner does: it enumerates candidate trajectories, scores them against explicit constraints, and cannot return a path it did not check. Running search downstream of learning gives the system a floor that does not depend on the neural network having generalized correctly.

The cost of getting behaviour prediction wrong is not abstract. When a Serve robot collided with a man using a mobility scooter in West Hollywood in September 2025, the company's own explanation was that "our safety system designed to predict pedestrians' intentions and yield right of way instead caused the robot to impede their way," after which "the robot came to a full stop in response to sensing a pedestrian in close proximity, which is considered its fail-safe state" ^[22]. The fail-safe fired correctly and made the outcome worse, because the man was accelerating around the robot when it stopped.

Training and the long tail

DoorDash says it trains "with behavior cloning and reinforcement learning on large, diverse data" ^[1]. Behaviour cloning is supervised learning on human demonstrations: given a sensor context, predict the action a competent driver took. It is cheap, captures normal driving well, and degrades badly out of distribution, because a cloned policy has never seen the states its own mistakes produce. Reinforcement learning, usually in simulation, is the standard corrective: the policy explores states no demonstrator generated and is scored against an explicit objective rather than imitation.

What makes delivery different from robotaxi development is the shape of the long tail. DoorDash names the gap directly: it is scaling RL for "the long tail of delivery-specific scenarios, like navigating around a child's bike left in a driveway, or safely approaching a front door when a dog is wandering by, scenarios that rarely appear in traditional AV datasets, but are daily realities for delivery" ^[1]. Autonomous-driving corpora are collected from road vehicles on roads. They hold enormous quantities of lane changes, unprotected turns and crossings, and almost nothing of driveways, porches, gates, hoses, lawn furniture or loose dogs. A delivery stack cannot borrow its way past that gap, which is why the training story here is closer to robot learning and sim-to-real transfer than to conventional autonomous driving. DoorDash also says it is building "a foundation driving model that seamlessly handles the transition between road driving, sidewalk navigation, and precise driveway maneuvering, all within a single inference pass" ^[1]. No results or deployment data for that foundation model have been published.

The last ten feet

DoorDash's engineers give the problem its name: "delivery poses a 'last ten feet' problem. Unlike ride-hailing, where consumers take themselves to and from vehicles, many consumers want food or groceries delivered to their door" ^[1]. A robotaxi's task ends at a legal curb. A delivery robot's hardest work starts there.

The unsolved parts are mundane and numerous: gates and latches, steps and stoops, lobby doors and lifts, apartment numbering, unlit paths, dogs, and a handoff that has to work when nobody comes out. Most operators sidestep the problem by requiring the customer to walk to the kerb and unlock a lid from a phone. That is a product decision dressed as a technical one, and it caps how much labour the robot actually removes. Pickup is no easier: as DoorDash notes, "different merchants want the robot to park in different places. And who loads the delivery?" ^[1] In current deployments, a human does.

Teleoperation and remote supervision

This is the least reported and most important fact about the sector. Nearly every commercially deployed delivery robot keeps a human in the loop, and several run with humans driving. What each operator discloses varies enormously, and some disclose nothing.

The best-documented case is Serve Robotics, and it became documented by accident. In September 2022 a Serve robot in Hollywood crossed under police tape and rolled through an active crime scene. Chief executive Ali Kashani told TechCrunch that the machine had not decided this: a remote human operator had. "The robot wouldn't have ever crossed (on its own)," he said; the operator stopped at the tape, then proceeded after bystanders lifted it and waved the robot through. "The judgment error here is that someone decided to actually keep crossing." The same reporting established company policy at the time: human operators remotely monitored and assisted at every intersection, and took control whenever the robot could not resolve an obstacle within 30 seconds ^[23]. Serve's 2025 annual report puts it in filing language: Level 4 robots "can operate without humans in the loop for periods of time," and "the fleet is monitored through mobile connectivity and video streaming by remote human supervisors who can assist robots when necessary" ^[24].

Coco Robotics never claimed otherwise. Co-founder Zach Rash told Marketplace in 2023 that "a lot of companies are really focused on autonomy and R&D... We started with remote drivers." Coco's pilots are employees rather than gig workers, driving robots from home using Xbox controllers; Rash said the company recruits gamers, trains them in simulators, and considers it a better job than gig delivery ^[25]. Trade coverage of the DoorDash partnership describes Coco's machines as remotely piloted, with operators taking control at busy intersections, construction zones and specific drop-off instructions ^[26].

Starship's disclosures contradict each other inside a single FAQ page. Under "Are Starship robots really autonomous?" the company answers: "Yes! Starship has been operating at Level 4 autonomy since 2018, meaning little human interaction during each journey and nobody controlling them remotely. Currently, Starship operates at over 99% autonomy... For safety reasons, we also make sure to have human remote assistants on standby, in case they're called upon to support." Two other answers on that same page say the robots "drive autonomously on pavements/sidewalks, but are monitored by humans who can take control at any time," and that they "are constantly monitored by a remote assistant" ^[19]. "Nobody controlling them remotely" and "constantly monitored by a remote assistant" are not the same claim, and a reader has no way to tell which describes the fleet. The "over 99% autonomy" figure is company-reported, carries no published methodology or as-of date, and is the sector's most-cited autonomy statistic.

Cartken states its dependency plainly, which is unusual: robots "can notify remote human operators in edge cases," operators have an "Instant Override" over LTE, and, most tellingly, "Robots pause at mapped crossings until given human approval" ^[14]. Avride says a "dedicated remote support team" can "control the robots remotely if an extraordinary situation arises" ^[21].

DoorDash's account of Dot is contested on the record. DoorDash Labs autonomy chief Ashu Rege told TechCrunch at the unveiling that Dots "cannot be remotely operated by humans" and that DoorDash believes teleoperation is not the right approach, describing a robot trained to wait and pull over until a field operator is dispatched ^[27]. DoorDash's own consumer page for Dot says the opposite: "If Dot encounters an issue, remote operators can step in to assist, and a local operations team is on hand to resolve anything that can't be managed remotely" ^[28]. The City of Fremont's permit splits the difference and is precise about the gradient: in Phase 1A, DoorDash "would only operate up to 3 robots for community engagement, testing and demonstration purposes, during which robots are chaperoned by local operators," and only an amended permit would later "allow up to 30 robots to operate autonomously within the Service Area and with the capability to be operated remotely as needed" ^[50]. Arizona law, meanwhile, makes remote human oversight a precondition rather than an option, as set out below.

Chinese operators are the most matter-of-fact: Zelos, Rino.ai and Neolix run remote monitoring centres in which a vehicle meeting an obstacle slows and pulls to the kerb automatically, and operators intervene only on unusual situations ^[7].

No operator found in researching this article publishes an operator-to-robot ratio, an intervention rate, or a disengagement report. Serve publishes daily active robots and daily supply hours, which bound the problem from one side but say nothing about how many humans were watching ^[17]. Every supervision figure in this article exists because a company volunteered it or a journalist extracted it after an incident. "Level 4" is used across the sector as a marketing badge with no published operational design domain attached to it. Until operators publish intervention data on the model of state autonomous-vehicle disengagement reporting, anyone comparing autonomy across these companies is comparing marketing copy, not measurements.

Operators

All scale figures below are company-reported and unaudited. Dates are the as-of dates the companies themselves attach.

Scroll sideways for more →

Regulation

Personal delivery device statutes

The United States has no federal sidewalk-robot rule. Instead, states have passed personal delivery device statutes, mostly since 2017, and at least 23 had done so by the end of 2022 ^[34]. They differ on exactly the parameters that determine which machines qualify: Georgia permits devices up to 500 pounds but caps sidewalk speed at 4 mph, while New Hampshire allows 10 mph but caps weight at 80 pounds ^[34]. A robot legal in one state can be categorically outside the definition in the next.

Arizona, the permissive case

Arizona is the most permissive major jurisdiction, and it is where Dot launched. Its first PDD law, HB 2422, was chaptered on 8 May 2018: it defined a personal delivery device as an electronically powered device operating primarily on sidewalks and crosswalks, at a maximum of 7 mph and weighing less than 200 pounds excluding cargo, and it repealed itself on 1 September 2020 ^[35]. The 2020 replacement rewrote both the definition and the operating rules, and the differences are decisive ^[36].

Under the current ARS 28-101(63), a personal delivery device is a device "manufactured for transporting cargo and goods" in the areas described by 28-1225 and "equipped with automated driving technology, including software and hardware, that enables the operation of the device with the remote support and supervision of a human" ^[37]. There is no weight limit anywhere in the definition. That absence is what lets a 350-pound machine qualify. Arizona keeps a separate 80-pound, 12 mph category, the personal mobile cargo carrying device, which is explicitly excluded from the PDD definition ^[37].

The operating provisions:

Scroll sideways for more →

Two features stand out. First, Arizona defines away the autonomy question: the definition requires "remote support and supervision of a human," and 28-1223 requires an agent capable of monitoring or exercising physical control. Human oversight is not a fallback in Arizona, it is a condition of legality. Second, the statute preempts most local rulemaking, which is why municipal arrangements such as Fremont's encroachment permit, in a state without a comparable PDD law, look nothing like Arizona's regime.

Cities that pushed back

San Francisco is the canonical restriction. After sidewalk robots appeared in 2016 and 2017, Supervisor Norman Yee introduced an outright ban, then, lacking the votes, amended it into a permit programme backed by Walk SF, Senior and Disability Action and Teamsters Joint Council 7 ^[38]. The ordinance the Board of Supervisors passed in December 2017 limited each company to three robots and the whole city to nine at any time, confined them to low-foot-traffic industrial areas, capped speed at 3 mph, required a human operator within 30 feet of each machine, and obliged permit holders to submit monthly GPS and photographic testing data ^[39]. It was in effect a testing licence rather than a commercial one.

Accessibility

The most substantive criticism of sidewalk robots is that they occupy a public accommodation disabled people depend on and cannot route around. In October 2019 the University of Pittsburgh paused its newly launched Starship programme after a doctoral student and wheelchair user, Emily Ackerman, described being trapped in the roadway on Forbes Avenue because a robot blocked the only accessible entrance to the sidewalk; another wheelchair user, Alisa Grishman, had been blocked earlier that month. Starship said it had reviewed the footage and concluded Ackerman was able to access the sidewalk ^[40].

Six years later the same failure mode recurred far more visibly. On 12 September 2025 Mark Chaney, a disability advocate with cerebral palsy, filmed a Serve robot repeatedly swerving into the path of his mobility scooter in West Hollywood before braking abruptly, causing a collision; the video passed 26 million views. The LA Times reported that the collision took place "within days of another incident where a Serve robot was recorded blocking an activated LAFD truck in Hollywood." Serve said its robots are designed and tested to navigate safely around mobility devices, "We regret when we do not live up to that," and blamed its pedestrian-intention prediction system ^[22]. Chaney asked the company to convene an accessibility council. These are not edge cases in the statistical sense: curb ramps are exactly where a sidewalk robot must wait to cross, and exactly the point of failure for a wheelchair user. No operator publishes accessibility incident data.

Economics

The economic case is that a two-ton car is the wrong tool for a two-pound burrito. Serve's Kashani has made the argument in those terms for years, adding that "a car has about 3,000 times more kinetic energy than one of our robots" ^[7]. Refraction AI's pitch was more specific: a machine cheap enough (roughly $5,000 per unit in 2019) that capital cost per delivery could undercut a courier ^[18].

Whether it does is almost entirely unpublished. No operator discloses an absolute cost per delivery. The one exception is a claimed differential: announcing its move into grocery on 4 June 2026, Heinla said "our robots deliver groceries at a cost $3-4 lower per delivery than traditional courier fulfilment," a company-reported figure with no methodology attached ^[48]. Serve, the only listed pure play, gives the closest thing to an audited answer and it is not flattering: revenue of $2.7 million against a net loss of $101.4 million for 2025, with a fleet exceeding 2,000 robots at year end ^[24]. In the first quarter of 2026 it reported $3.0 million of revenue from 812 daily active robots and 10,295 daily supply hours, said gross margin percentage had "improved meaningfully" over the prior quarter, and told investors its focus had shifted from fleet expansion to revenue per robot and per operating hour ^[17]. Coco has never published unit economics; DoorDash has published neither a Dot fleet size nor a delivery count.

Hardware cost is visibly collapsing in China. According to The Wire China, a vehicle that cost more than 1 million yuan a few years ago can now sell for under 20,000 yuan, with Zelos at 19,800 yuan, Cainiao launching one at 16,800 yuan, and Neolix offering instalment plans from an 888 yuan deposit. A ZTO pickup-station owner said his sub-20,000-yuan vehicle would pay for itself in two months, against "about 3,000 yuan per month" for a rented truck and "around 8,000 yuan per month" for a driver, and Rino.ai claims a 30 to 50 percent last-mile cost reduction on fixed courier routes. The same report quotes Whale Dynamic's chief executive calling the price war unsustainable: "The story [ADV makers] sell is that they're building volume fast to attract capital and go public, but the reality is the financing is showy, the profits aren't there" ^[7].

Two structural costs are systematically understated. The first is remote supervision, which The Wire China notes "reassures U.S. regulators and the public but adds to costs" ^[7]. The second is field operations: chargers, retrieval trucks, cleaning, repair and the humans who load merchant handoffs and unstick robots. Neither appears in any published per-delivery figure. Market-size and growth-rate projections for this sector circulate widely; every such figure encountered in researching this article traced to a search-optimised content mill with no stated methodology, and none is cited here.

Criticism

Sidewalk obstruction and accessibility. Covered under Regulation above; it is the criticism with the strongest documentary record and the weakest industry response ^[22][40].

Labour. Organised labour has opposed delivery automation in San Francisco at both ends of the technology. Teamsters Joint Council 7 backed the 2017 sidewalk-robot ordinance ^[38]. In November 2025 Teamsters Local 665 appealed DoorDash's planned drone testing laboratory in the Mission District; principal officer Tony Delorio said "this PDR property is meant for blue-collar jobs, but DoorDash is using it to develop technology designed to destroy jobs." The Board of Appeals denied the appeal on 19 November 2025, and the union filed a second appeal on 26 November on the ground that only three of five board members had been present ^[41]. The city then legislated: on 16 December 2025 the Board of Supervisors adopted File 251116, sponsored by Supervisors Jackie Fielder and Shamann Walton, imposing 18 months of interim zoning controls that require conditional use authorization for outdoor development and engineering laboratories in the PDR-1-G district. The mayor approved it on 23 December 2025 ^[51]. DoorDash's framing is that Dashers "will continue to complete the vast majority of our millions of daily deliveries" while autonomy handles lower-value trips ^[2]. In China the framing is less hedged: JD Logistics said in October 2025 it intends to procure three million robots, one million autonomous vehicles and 100,000 drones over five years ^[33].

Vandalism and theft. Robots on public sidewalks are unattended property. Futurism's April 2026 survey documented Avride robots kicked over in Philadelphia, Starship robots in Sheffield spray-painted with "off our streets," an influencer collective filming itself thrashing a Coco robot in Los Angeles, and one machine cracked open and abandoned; it reported around 1,600 of 80,000 Kiwibot campus deliveries in Berkeley involving a vandalism incident, against a per-robot cost of about $2,500 ^[42].

Emergency services. The 2022 Serve crime-scene crossing ^[23], the 2025 LAFD truck obstruction ^[22], and the June 2026 incident in which a Dot entered an active SWAT operation in Chandler, Arizona, stayed through a flashbang detonation and had to be piloted manually onto a DoorDash box truck afterwards ^[43] form a pattern: these machines have no reliable model of an emergency scene. DoorDash said "our robot behaved as designed, stopping and waiting safely while authorities managed the scene" ^[43], which is accurate and also concedes the point, since what authorities wanted was for it to leave.

Privacy. A sidewalk robot is a continuously recording, mobile camera array in public space. Starship says that under remote assistant control the image feed is lower resolution and obfuscated to conceal identities, and that high-resolution imagery captured in autonomous mode is retained only briefly ^[19]. DoorDash says Dot does not use facial recognition and that limited sensor data may improve the driving system but is never shared with partners and goes to authorities only if legally required ^[28]. These are unaudited policy statements, not verifiable technical constraints.

Overstated autonomy. As set out above, operational reality in this sector generally surfaces only through an accident, a regulatory filing or a permit condition, and rarely matches the marketing.

See also

• Drone delivery
• DoorDash Dot
• DoorDash Labs
• Autonomous Delivery Platform
• Autonomous mobile robot
• Teleoperation
• Autonomous vehicle
• Robot.com

References

1. ^Stanley Tang and Ashu Rege, "Engineering Autonomy for Local Commerce: Building Dot and the Autonomous Delivery Platform", DoorDash engineering blog, 30 September 2025. careersatdoordash.com/...onomous-delivery-platform
2. ^"DoorDash Unveils Dot, the Delivery Robot Powered by its Autonomous Delivery Platform to Accelerate Local Commerce", DoorDash newsroom, 30 September 2025. about.doordash.com/...doordash-unveils-dot
3. ^Kristin Musulin, "NHTSA grants first driverless vehicle exemption", Smart Cities Dive, 7 February 2020. smartcitiesdive.com/...571868
4. ^"About" (company history and annual figures), Starship Technologies, accessed 1 August 2026. starship.xyz/about
5. ^Lora Kolodny, "Marble and Yelp Eat24 start robot food delivery in San Francisco", TechCrunch, 12 April 2017. techcrunch.com/...t-food-delivery-in-san-francisco
6. ^Sean Scott, "Meet Scout", About Amazon, 23 January 2019. aboutamazon.com/...meet-scout
7. ^Peiyue Wu, "China's Delivery Revolution", The Wire China, 26 October 2025. thewirechina.com/...na-automated-delivery-vehicles
8. ^"Caterpillar acquires delivery robot maker Marble", Robotics and Automation News, 24 June 2020. roboticsandautomationnews.com/...33385
9. ^"Amazon ends testing of Scout delivery robots", The Robot Report, 6 October 2022. therobotreport.com/...ing-of-scout-delivery-robots
10. ^Kirsten Korosec, "Uber spins out delivery robot startup as Serve Robotics", TechCrunch, 2 March 2021. techcrunch.com/...-robot-startup-as-serve-robotics
11. ^"Our Company" (timeline), Serve Robotics, accessed 1 August 2026. serverobotics.com/company
12. ^"Company", Nuro, accessed 1 August 2026. nuro.ai/company
13. ^Rebecca Bellan, "Nuro pivots to license self-driving tech to carmakers, mobility companies", TechCrunch, 11 September 2024. techcrunch.com/...-to-carmakers-mobility-companies
14. ^Cartken company website, accessed 1 August 2026. cartken.com
15. ^Kirsten Korosec, "From burritos to biotech: How robotics startup Cartken found its AV niche", TechCrunch, 3 July 2024. techcrunch.com/...artup-cartken-found-its-av-niche
16. ^"Kiwibot relaunches as Robot.com with 1.7 million real-world robot tasks completed", Robotics and Automation News, 31 October 2025. roboticsandautomationnews.com/...96107
17. ^"Serve Robotics Announces First Quarter 2026 Results with 3X Sequential Revenue Growth", Serve Robotics investor relations, 7 May 2026. investors.serverobotics.com/...ter-2026-results-3x
18. ^Kirsten Korosec, "This new autonomous startup has designed its delivery robot to conquer winter", TechCrunch, 11 July 2019. techcrunch.com/...delivery-robot-to-conquer-winter
19. ^"FAQ", Starship Technologies, accessed 1 August 2026. starship.xyz/faq
20. ^"Our Robots" (sensor suite, battery, capacity) and "Company" (scale figures, "99% autonomous and operate at Level 4"), Starship Technologies, accessed 1 August 2026. starship.xyz/our-robots and starship.xyz/company
21. ^"Delivery Robot", Avride, accessed 1 August 2026. avride.ai/robot
22. ^Clara Harter, "A delivery robot collided with a disabled man on L.A. street. The aftermath is getting ugly", Los Angeles Times, 25 September 2025. latimes.com/...elchair-sparks-accessibility-debate
23. ^Kirsten Korosec, "This robot crossed a line it shouldn't have because humans told it to", TechCrunch, 16 September 2022. techcrunch.com/...otics-sidewalk-robot-crime-scene
24. ^Serve Robotics Inc., Form 10-K for the fiscal year ended 31 December 2025, EDGAR accession 0001832483-26-000010, filed 12 March 2026 (accepted after hours on 11 March); read via the StockTitan rendering. stocktitan.net/...files-annual-report-05d372ace5b4
25. ^Kai Ryssdal and Maria Hollenhorst, "The robot delivering your meal might have a human driver", Marketplace, 27 June 2023. marketplace.org/...robot-food-delivery-autonomy
26. ^"DoorDash and Coco deploy robot fleet with remote operators for urban delivery", R&D World, 11 April 2025. rdworldonline.com/...-operators-for-urban-delivery
27. ^Maxwell Zeff, "DoorDash unveils Dot, its autonomous robot built to deliver your food", TechCrunch, 30 September 2025. techcrunch.com/...robot-built-to-deliver-your-food
28. ^"Meet Dot" (product page and FAQ), DoorDash, accessed 1 August 2026. about.doordash.com/...dot
29. ^"DoorDash and Coco Expand Global Partnership with U.S. Sidewalk Robot Delivery Launch", DoorDash investor relations, 10 April 2025. ir.doordash.com/...default
30. ^"DoorDash and Coco Robotics Expand Partnership to Miami", PR Newswire, 13 November 2025. prnewswire.com/...oss-grocery-and-retail-302614124
31. ^"Avride secures strategic investment and other commitments of up to $375 million, backed by Uber and Nebius", Nebius newsroom, 22 October 2025. nebius.com/...75-million-backed-by-uber-and-nebius
32. ^"Neolix Completes Delivery of 10,000 Autonomous Vehicles", Neolix, 23 September 2025. neolix.ai/...10000-vehicles-delivery
33. ^"JD Logistics unveils five-year plan to deploy millions of robots, autonomous vehicles and drones", TechNode, 27 October 2025. technode.com/...ots-autonomous-vehicles-and-drones
34. ^Max Garland, "Why Delivery Robots Face a Regulatory 'Nightmare'", Supply Chain Dive, 26 April 2023. supplychaindive.com/...648303
35. ^"HB 2422 House Bill Summary" (chaptered 8 May 2018), Arizona State Legislature. azleg.gov/...H.HB2422_05-08-18_CHAPTERED.DOCX
36. ^"Personal delivery device", Arizona Bike Law (legislative history of HB 2422 and SB 1304). azbikelaw.org/personal-delivery-device
37. ^Arizona Revised Statutes Title 28, sections 28-101, 28-1221, 28-1222, 28-1223, 28-1224, 28-1225, 28-1226, 28-1227 and 28-1228, Arizona State Legislature. azleg.gov/...00101
38. ^"San Francisco Supervisor Seeks Permit Program For Delivery Robots", CBS News Bay Area, 18 October 2017. cbsnews.com/...ermit-program-supervisor-norman-yee
39. ^Katie Pyzyk, "San Francisco tightens regulations on delivery robots", Smart Cities Dive, 12 December 2017. smartcitiesdive.com/...512793
40. ^Chris Albrecht, "Starship's Robots Pulled From University of Pittsburgh After Accessibility Incident", The Spoon, 24 October 2019. thespoon.tech/...urgh-after-accessibility-incident
41. ^Kelly Waldron, "DoorDash wants to test drones in the Mission. Teamsters are pushing back.", Mission Local, 19 November 2025. missionlocal.org/...e-testing-mission-pdr-district
42. ^Joe Wilkins, "Delivery Robot Companies in Trouble as Bot Become Targets for Vandalism", Futurism, 2 April 2026. futurism.com/...delivery-robot-vandalism-problem
43. ^Frank Landymore, "DoorDash Delivery Robot Invades SWAT Scene, Won't Leave as Cops Flashbang Resident", Futurism, 24 June 2026. futurism.com/...delivery-robot-swat-scene
44. ^Steve Mollman, "DoorDash's drone delivery approval" (referring to Dot as "its autonomous sidewalk-delivery robot"), Newsweek, 30 July 2026. newsweek.com/...orker-skills-faa-approval-12265674
45. ^Ishveena Singh, "DoorDash Air drone delivery launch", DroneDJ, 31 July 2026. dronedj.com/...doordash-air-drone-delivery-launch
46. ^Joseph Ofonagoro, "DoorDash Air: FAA drone delivery approval", TechRepublic, 30 July 2026. techrepublic.com/...oordash-air-faa-drone-delivery
47. ^Kirsten Korosec, "DoorDash is building its own drone delivery business", TechCrunch, 29 July 2026. techcrunch.com/...-its-own-drone-delivery-business
48. ^"Starship Technologies Doubles Down on Grocery Delivery in EU and US Markets", Starship Technologies press release, 4 June 2026. starship.xyz/...chnologies-doubles-down-on-grocery
49. ^Jan Schlain, "Robot Farewell", Ann Arbor Observer, 19 November 2021. annarborobserver.com/robot-farewell
50. ^"DoorDash Delivery Robots", City of Fremont, archived capture of 21 April 2026 (the live page is unreachable from this environment). web.archive.org/...doordash-delivery-robots
51. ^"Interim Zoning Controls: Conditional Use Authorization for Outdoor Laboratory Uses in PDR-1-G", San Francisco Board of Supervisors File 251116, adopted 16 December 2025, Enactment 600-25. sfgov.legistar.com/LegislationDetail
52. ^Rebecca Szkutak, "Sam Altman-backed Coco Robotics raises $80M", TechCrunch, 11 June 2025. techcrunch.com/...-backed-coco-robotics-raises-80m