Industries / Health / Healthcare / Pharmacy
Automation
Machines that count, store, pick, pack and hand out medicines, in pharmacies, on hospital wards and in central filling plants.
7
Milestones achieved, from 1969 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1969
A photocell counts pills
Harry Roseberg, a California salesman, had the idea in 1962 while watching a pharmacist count by hand. His machine lines pills up on a turntable, counts them as they pass a photoelectric cell and drops them into a box. His firm leased 75 counters to Los Angeles County drugstores for two years of testing, then signed McKesson & Robbins to market it nationally; by December 1969 Time reported 1,724 leases at $17.50 a month.
Credit
Harry Roseberg and Irving Zeiger, RX COUNT Corp., Hawthorne, California
1990
Drug cabinets reach wards
Pyxis Corp. of San Diego began selling the Med Station in 1990: a locked metal cabinet on the ward whose drawers release a drug when a nurse keys in the order, and which records who took what. The Los Angeles Times reported sales of $300,000 in 1990 and $13.4 million in 1991, and a Long Beach hospital with 24 units said its records of controlled drugs had become far more accurate. In a controlled before-and-after study of two intensive care units at Grenoble University Hospital, the unit given a cabinet cut its error rate from 20.4% to 13.5% of opportunities, but errors that caused harm did not fall.
LessonFewer slips at the cabinet did not mean fewer harmful errors.
Credit
Ron Taylor, Pyxis Corp.; Claire Chapuis and colleagues (Critical Care Medicine, 2010)
1994
A plant fills by mail
The US Department of Veterans Affairs had pooled mail prescriptions from several hospitals on a small scale in the 1970s and 1980s. In 1994 its Consolidated Mail Outpatient Pharmacy (CMOP) at Leavenworth, Kansas, began filling them in high volume on an integrated, automated dispensing line, and six more plants followed. Auditors from the VA's Office of Inspector General found that about 9.3 million of 9.4 million packages in October 2015 arrived within the 10-day goal; the VA said the plants handled about 80% of its outpatient prescriptions.
Credit
US Department of Veterans Affairs; audit by the VA Office of Inspector General (2016)
2013
A pack-picking robot, measured
An original-pack robot stores sealed manufacturer's packs and fetches the one ordered, which suits countries where patients get whole packs rather than counted tablets. Researchers collected data over six weeks at an NHS hospital pharmacy in Wales, before and after one was installed. Items dispensed per person per hour rose from 9.20 to 13.17, and dispensing errors caught before they left the pharmacy, as reported by staff, fell from 0.64% to 0.28% of items.
LessonErrors still rose with workload after the robot arrived.
Credit
K. L. James, C. Whittlesea and colleagues, University of Bath (International Journal of Pharmacy Practice, 2013)
2023
Robot hubs paused, then restarted
Walgreens, a US chain, began opening robot-run centers in 2021 that fill prescriptions for many stores and truck them back for pickup. It paused expansion in 2023 to collect feedback and improve the existing sites, and said in May 2025 that it was ready to expand again. By Walgreens' figures, reported by CNBC, 11 centers then served 4,800 stores, filled about 16 million prescriptions a month and handled 40% of the volume of the pharmacies they supported.
LessonCentral filling took two years of fixes, guided by the stores it served, before it could grow again.
Credit
Walgreens; reported by CNBC
2024
A kiosk maker goes bankrupt
MedAvail built the MedCenter, a kiosk in a clinic that hands over a prescription medicine after the patient talks to a remote pharmacist by video. In January 2023 it said it would close its own SpotRx pharmacies, which ran on the kiosks, cut about 75% of its staff and live on selling the machine, a business it expected to bring in $1.2 to $1.4 million in 2022. On 2 February 2024 it filed for Chapter 7 bankruptcy, the US form that winds a company up.
LessonA kiosk that worked did not make a pharmacy that paid.
Credit
MedAvail Holdings (its filings with the US Securities and Exchange Commission)
2025
Hubs may serve other owners
In hub-and-spoke dispensing a local "spoke" pharmacy takes the prescription and a large "hub", often automated, assembles the medicines and sends them back. UK law allowed this only between pharmacies of one company, which shut independent pharmacies out of automated hubs. Regulations made on 25 June 2025 and in force from 1 October 2025 let a hub work for pharmacies owned by others.
Credit
UK Department of Health and Social Care; Department of Health, Northern Ireland
Watch for
Humanoids staff Beijing pharmacies
Galbot says its G1 robots find medicines on open shelves, pick and pack them and put each order in a pickup cabinet, around the clock, in a pharmacy of about 70 square meters that stocks some 5,000 products. It reports a grasp success rate of about 99.5%. In July 2025 a co-founder said the robots ran in more than 10 Beijing pharmacies and that the firm planned 100 such stores nationwide by the end of that year.
Achieved whensomeone outside the company publishes, for a named store, how many orders the robot filled without staff help and how often it picked the wrong item.
Claimed by
Galbot (co-founder Zhang Zhizheng), July 2025, and through China's state broadcaster CGTN, March 2026
Kiosks in doctors' offices
Amazon Pharmacy says kiosks in One Medical clinics across greater Los Angeles hand patients their medicines within minutes of an appointment from December 2025, after an Amazon pharmacist reviews each order, with more offices to follow.
Achieved whena regulator or outside study reports how many prescriptions the kiosks filled and how many were wrong or never collected.
Claimed by
Amazon, October 2025
Open
01
Shared test for pill counters
A public set of phone photos of dummy tablets (candy, beads, empty capsules, vitamin supplements) lying in a counting tray, each with its true count, and a script that scores any counting app or AI model against it. The photos cover touching, overlapping, clear and on-edge tablets under several kinds of light.
The briefCredit · unclaimed02
Look-alike carton picking test
A public test in which a low-cost robot arm takes a named carton from a tightly packed shelf row of look-alike cartons, confirms it by scanning its barcode and sets it in a tray, then puts returned cartons back. All cartons are empty dummies printed with made-up names and codes.
The briefCredit · unclaimed
Open 01 · the brief
Shared test for pill counters
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of phone photos of dummy tablets (candy, beads, empty capsules, vitamin supplements) lying in a counting tray, each with its true count, and a script that scores any counting app or AI model against it. The photos cover touching, overlapping, clear and on-edge tablets under several kinds of light.
Why it is still open
The one outside evaluation of a counting app found, by pharmacists at a Tokyo hospital in 2025, tested a single app on three kinds of supplement. It found that clear tablets, overlap and orientation caused miscounts, and its authors wrote that no earlier study had evaluated these apps. A 2026 paper in JMIR AI by Liu and colleagues built its own counting model and scored it on a public set of 152 phone photos, with a mean error of 0.167 pills, but it did not test the apps pharmacies use or counts of the size a pharmacy fills. Three searches of Hugging Face for pill-counting datasets returned nothing.
Built with what exists
a phone and a counting tray; CVAT or Label Studio for labeling; OpenCV and Ultralytics YOLO as baselines; a vision-language model such as Gemini or Claude; the counting apps on sale, including the one in the Tokyo study; Zenodo or Hugging Face for release.
First step
Photograph three candy shapes at counts of 30, 60 and 90, flat and overlapping, run one app and one model on them, and post the photos and results.
Done looks like
At least 2,000 labeled photos of 20 or more dummy shapes, with counts from 1 to 120, published with a DOI, and a table of the exact-count rate and size of miscounts for at least three apps and three models, split by condition, that anyone can rerun.
Open 02 · the brief
Look-alike carton picking test
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test in which a low-cost robot arm takes a named carton from a tightly packed shelf row of look-alike cartons, confirms it by scanning its barcode and sets it in a tray, then puts returned cartons back. All cartons are empty dummies printed with made-up names and codes.
Why it is still open
Galbot's 99.5% grasp rate for its pharmacy robots is the company's own figure, with no public test behind it. The largest public robot recording of this task found on Hugging Face holds 102 episodes of picking and placing a medicine box, posted in September 2026 with no success tally, and a 2024 paper on grasping in a "smart pharmacy" reports its own lab experiments. A shelf-picking benchmark that anyone can rebuild, with a fixed shelf, stencils for placing items and a scoring protocol, has existed since 2016 (the ACRV Picking Benchmark), but its 42 objects are household items, not look-alike medicine packs. Searches found no shared benchmark for medicine cartons.
Built with what exists
an SO-101 arm and Hugging Face's LeRobot software (see the Note "No download"); a robot model with open weights, such as Runway Praxis-1 (see the Note "Open robot model weights"); the ACRV benchmark's stencil and scoring rules as a pattern; two USB cameras; the zxing-cpp barcode library for the GS1 DataMatrix codes printed on medicine packs; a color printer and card for the dummy cartons.
First step
Print 10 dummy cartons in two look-alike families, record 50 remote-controlled picks with LeRobot and post the dataset.
Done looks like
50 dummy cartons in 10 look-alike families, 500 filmed retrievals with the shelf reshuffled between runs, and a published tally of right carton, wrong carton, drops and crushed cartons, with and without the barcode check. Video, code, data and the carton print files are released.