Delivery Robots Compared: Sidewalk vs Hospitality vs Hospital

Three categories of delivery robot are usually discussed as one market. They should not be. Sidewalk robots, restaurant tray-runners and hospital logistics units face different regulation, different economics and — most importantly — replace very different kinds of work. That last difference decides almost everything.

A warning before the numbers: this sector's pricing is deliberately opaque. Bear Robotics, Diligent and Relay all publish "request a quote" and nothing else. Widely-circulated figures for Diligent's Moxi trace to AI-generated content sites and are excluded here. Where a real number does not exist publicly, this article says so.

The pricing that is actually published

Hospitality purchase prices come from RobotLAB, an established US distributor — reseller listings, not manufacturer MSRP. The TUG figures carry an important date caveat: they are from 2008. A 2015 article circulates with identical numbers, which suggests recycling rather than a fresh quote. No current hospital-robot pricing could be verified.

On the widely-repeated "$1,000–1,500/month" figure for restaurant robots: it did not hold up. Bear publishes no pricing at all. An independently reported 2025 deployment in Virginia paid $1,000/month for two units — about $500 each, consistent with Pudu's verified $335–$395. Realistic current RaaS is roughly $335–$500 per unit per month; higher figures appear to be early-market or bundled-service pricing.

Sidewalk: the audited numbers contradict the marketing

Serve Robotics is the only pure-play sidewalk operator with public financials, and they are sobering. FY2025: $2.7M revenue against a $15.4M gross loss, a $101.4M net loss, with $260M cash on hand.

Serve markets "under $1 per delivery at scale versus $8–10 status quo." That is a forward-looking target, not a realized cost — and the company discloses no actual cost per delivery, revenue per delivery, or deliveries per robot.

Two derived observations, both from Serve's own disclosures. First, utilization: 2,000 robots deployed against 547 daily active robots in Q4 2025 — roughly 27% of the fleet works on an average day, a figure the company does not address. Second, revenue per daily-active robot appears to have *fallen* as the fleet scaled.

Starship claims per-delivery profitability in select markets but publishes no dollar figure and is private and unaudited. Coco, Kiwibot and Avride disclose no unit economics at all. For a purchase decision, sidewalk cost-per-delivery claims are currently unfalsifiable — with Serve's filings the sole exception, and those contradict the headline.

A strategic signal worth weighing: Starship is withdrawing 1,200+ robots from 60+ US campuses, with its CEO describing campuses as "testing ranges." Anyone evaluating a campus deployment should ask the vendor to address that directly.

The regulatory and accessibility exposure is real

At least 23 states had personal delivery device laws by end-2022 — treat that as a floor; the authoritative current count could not be verified. The variation is extreme. Pennsylvania classifies these robots as "pedestrians" and permits 550 lb units. Virginia requires $100,000 liability insurance. San Francisco permits R&D testing only, capped at nine devices citywide. Toronto banned them from sidewalks outright in 2021; Knoxville prohibited them city-wide in 2024.

Note also that the weight caps in several state laws were shaped by Amazon and FedEx lobbying tailored to their own robots' dimensions rather than by safety research.

The accessibility record is the best-documented failure evidence in the sector. A wheelchair-using PhD student at the University of Pittsburgh was trapped in the roadway when a robot blocked the curb cut, her only route back to the sidewalk. In West Hollywood in 2025, a man with cerebral palsy on a mobility scooter was repeatedly cut off until he rear-ended the robot — and Serve conceded fault, stating its yielding system "instead caused the robot to impede their way." Peer-reviewed work at CHI 2024 found 14 of 15 participants with mobility disabilities reacted negatively, with curb cuts the universal flashpoint.

One more disclosure buyers rarely consider: internal emails show Serve provided robot camera footage to the LAPD — proactively in March 2023 before any subpoena.

Hospitality: the pilot-to-scale gap

This category has the friendliest regulatory position by far — no permits, no municipal approval, no PDD statutes. Constraints are health-code and ordinary workplace safety. That is a genuine structural advantage.

The problem is operational, and the Chili's case documents it precisely. In a 10-store pilot, 82% of guests said the robot improved their experience. Scaled to 61 stores, 58% of guests surveyed said it did not — the robot moved too slowly and obstructed staff. Brinker paused the program, its CEO citing projects without "a line of sight to a return on the business."

An operator supplied the mechanism: robots "were working very well during the pandemic when we had fewer customers and more spacing between tables." The 2020–22 adoption wave happened under artificially sparse conditions. Density is what these machines handle worst, and density is what a profitable dining room looks like.

Two further constraints that vendor ROI models tend to omit:

• A human still loads and a human still finishes the table. Every tray-based model here — BellaBot, Servi, Keenon, Matradee — is a runner assist. None plates or clears autonomously. Independent reporting documents staff *following* the robot during service, which adds labor.
• Layout must be static. Robots get confused when tables are moved or combined, which is a normal daily occurrence. BellaBot is rated to a ≤5° climb angle — no steps, no ramps, and carpet is a documented problem.

Among public vendors, Richtech Robotics posted FY2025 revenue of $5.0M with net loss widening to $15.8M, and Q1 FY2026 revenue *declining* ~9% year over year.

Where the evidence is genuinely positive: one Michigan restaurant cut staffing from five or six to three for equivalent volume, independently reported. Multiple operators cite real physical relief from carrying hot dishes. The pilot results were not fabricated — they just did not survive scale.

Hospital: the only defensible ROI case

El Camino Hospital, running roughly 19–20 TUGs across three buildings for about a decade, reports a 12.6 FTE offset, ~$650,000/year in wages and benefits, and costs 40% below human labor for the same tasks — with about 80% of all material deliveries automated. In the first year alone: 12,700+ delivery hours displaced, roughly one third of them pulling carts over 400 lbs.

That last detail is the whole argument. Hospital robots replace heavy, injury-prone, distance-intensive work. Restaurant robots replace carrying a tray thirty feet. This asymmetry explains why hospital ROI evidence is strong and hospitality ROI evidence is not.

The figures are customer-reported in an industry whitepaper, so treat them as a favorable case — and note they are now eight years old, which is itself a comment on how little hard data this sector produces.

Elevators are the least transparent line item

Three architectures, three very different install bills:

• Aethon requires the hospital's own elevator service team to wire a ReadyElevator Interface to the controller — real capital work, with commissioning cost and timeline unpublished. It does handle fire and safety signals, which matters for code.
• Relay markets proprietary compatibility across the major elevator makers explicitly as riding elevators "without the need for expensive elevator integration" — a direct jab at the above.
• Moxi installs nothing: a robotic arm presses the button and badge-scans on secure floors, with 110,000 cumulative autonomous rides logged.

Tellingly, Diligent built busy-bay avoidance into Moxi — elevator contention is a real problem the vendors quietly design around. Get commissioning cost in writing before signing.

Two cautions specific to hospitals

Security. The JekyllBot:5 vulnerabilities disclosed in Aethon TUG systems in April 2022 included CVE-2022-1070 at CVSS 9.8 — an unauthenticated attacker could take full control of robots, operate doors and elevators, hijack integrated cameras, and pivot malware across the hospital network. Patched in firmware v24; all prior versions were vulnerable. No real-world exploitation is known.

Workflow friction. A BMC Nursing scoping review found a robot "was hardly tolerated" on an acute internal medicine ward, with nurses anxious when robots stopped in corridors for extended periods. A 2026 observational study across 53 hours of hospital units concluded these robots are "presently best suited for auxiliary and background roles." Acceptance varied sharply by unit — obstetrics integrated well, high-stress units did not.

How to choose

Ask whether the robot replaces heavy work or light work. That single question predicts the outcome better than any spec. Hospital logistics passes it; table-running mostly does not.

Budget the human loader as permanent, not transitional. Any vendor model omitting it is unsound.

Demand utilization data from a busy room at peak — every vendor reference customer is, by selection, a sparse-environment case.

Match the risk type to your tolerance: sidewalk risk is external and political (permits, bans, accessibility litigation); hospital risk is internal and procedural (infection control, pharmacy governance, cybersecurity review); hospitality has neither, which is its main advantage.

For category background see delivery robots and our delivery robot cost and ROI guide.

One reasoning trap to avoid: Eatsa, Cafe X and Zume were food-preparation startups. Their failures are routinely cited as evidence against table-delivery robots. That inference does not hold.

Sources

• Serve Robotics FY2025 results — audited revenue, gross loss, fleet KPIs
• Serve Robotics 8-K summary — the "<$1 per delivery" target and Adjusted EBITDA
• Richtech Robotics FY2025 10-K — revenue, RaaS contracts, widening losses
• Restaurant Dive — Brinker pauses Chili's robot program — 61-store pause and CEO rationale
• CBS/AP — robot waiters — 58% at-scale disapproval, unit cost, positive Michigan case
• Interesting Engineering — restaurant operators on robot servers — load/unload burden, table-moving confusion, the density insight
• Robotics Business Review — Robots in Hospitals whitepaper — El Camino FTE offset, savings, cart weights
• Machine Design — Robots making rounds — the only verified TUG pricing (2008)
• The Robot Report — Moxi elevator milestone — arm-based elevator operation, busy-bay avoidance
• Aethon ReadyElevator — controller-wired interface, fire signal handling
• HIPAA Journal — JekyllBot:5 — CVE set and attack surface
• Han et al., CHI 2024 — Co-designing Accessible Public Robots — 14 of 15 participants negative
• Futurism — West Hollywood incident — Serve's admission of fault
• 404 Media — Serve footage to LAPD — pre-subpoena disclosure
• Supply Chain Dive — delivery robot legislation — 23-state count, vendor-shaped weight caps
• Ohneberg et al., BMC Nursing 2023 — ward tolerance and corridor congestion
• Georgadarellis et al., JMIR 2026 — 53-hour observational study, "auxiliary and background roles"
• FoodService Director — Starship exits higher ed — 1,200+ robots withdrawn