Industries / Industrial equipment and services / Robotics
Safety standards
Safety standards are the rules, written by standards bodies and governments, for how robots must be built, guarded and used so that they do not hurt the people who work near them.
7
Milestones achieved, from 1979 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1979
The first deaths
On 25 January 1979 Robert Williams, 25, was killed at Ford's casting plant in Flat Rock, Michigan, when a machine that fetched castings from a five-storey storage rack struck him after he climbed the rack to retrieve parts by hand; in 1983 a jury awarded his family $10 million against the machine's maker, Litton. In Japan, a manual written for the labour ministry by the Japan Industrial Safety and Health Association (JISHA) says the country's first death caused by an industrial robot came in 1981. In December 1984 the US National Institute for Occupational Safety and Health (NIOSH) described a die-cast operator killed that July between the back of a hydraulic robot and a steel pole after he climbed over a safety rail, and asked for interlocked gates, sensors that stop the robot when someone crosses the barrier, and clearance around every moving part.
LessonNIOSH wrote that workers recognise the danger of the arm's reach but not of the robot's other moving parts, and that "a robot often requires more operational space than does a human worker.".
Credit
NIOSH; JISHA
1983
Japan fences robots in
After the 1981 death, Japan added robot articles to its Industrial Safety and Health Regulations. Article 150-4 requires a fence or enclosure wherever contact with a running robot could endanger a worker, while a 1983 Ministry of Labour notice, in force from 1 July 1983, excluded machines whose drive motors are each rated at 80 watts or less. On 24 December 2013 the ministry ruled that robots with larger motors may also work beside people without a fence if the employer's risk assessment finds no danger, or if the robot was built and installed to ISO 10218, the international robot safety standard.
LessonA fixed power line was simple to apply but did not measure danger; JISHA's manual says that once risk assessments became mandatory in 2006, robots sharing space with people were in practice severely limited until the 2013 ruling.
Credit
Ministry of Labour, later Ministry of Health, Labour and Welfare
2016
Rules for side-by-side work
A collaborative robot, or cobot, is one built to work in the same space as a person without a fence. In February 2016 the International Organization for Standardization (ISO) published ISO/TS 15066, a technical specification that adds requirements for collaborative industrial robot systems to ISO 10218. Its tables set limits on the force and pressure a robot may put on each region of the body when it touches a person, which have to be measured on the finished cell, not the bare arm.
Credit
ISO technical committee 299, Robotics
2020
Mobile robots get standards
In late 2020 RIA published ANSI/RIA R15.08-1, which it called the first national safety standard for industrial mobile robots, robots that drive themselves around factories and warehouses; it covers the design of the robot. In October 2023 A3 added R15.08-2, covering how an integrator fits a robot or a fleet into a site; until then, A3 said, integrators "had only general safety requirements for industrial machinery."
Credit
Robotic Industries Association and A3 (Carole Franklin, director of standards)
2023
Robot deaths counted
A NIOSH study of the US Census of Fatal Occupational Injuries found 41 robot-related deaths at work from 1992 to 2017; 83% involved fixed robots, 78% were people struck by a robot moving under its own power, and many happened during maintenance. A 2024 study of the US Occupational Safety and Health Administration's (OSHA) severe injury reports found 77 robot accidents from 2015 to 2022: injuries from fixed robots were mainly finger amputations and fractures of the head and torso, and those from mobile robots mainly fractures of the legs and feet.
LessonSanders and colleagues found the reports' free-text accounts too thin and called for more structured, detailed narratives.
Credit
Linda Layne (NIOSH); Nathan Sanders, Ece Şener and Karen Chen (North Carolina State University and University of Leeds)
2023
EU rules for learning machines
The European Union's Machinery Regulation, (EU) 2023/1230, adopted on 14 June 2023, replaces the 2006 Machinery Directive. It requires an outside body to check safety components, and machines with built-in systems, whose behaviour evolves through machine learning while they perform safety functions. It also requires autonomous mobile machines, where relevant, to have a remote supervisory function that can only stop, start or move them to a safe state. EUR-Lex lists it as applying from 20 January 2027; the original text said 14 January 2027, and later corrections were published.
Credit
European Parliament and Council of the European Union
2025
The core standard rewritten
ISO published new editions of ISO 10218 Part 1, for the robot itself, and Part 2, for robot applications and cells, in 2025: ISO's pages give February 2025, while A3 says January. A3 reports that Part 1 grew from 50 to 95 pages and Part 2 from 72 to 223, that the number of listed safety functions rose from single digits to about three dozen, that cybersecurity requirements were added, and that the collaborative content of ISO/TS 15066 was folded into Part 2. The United States adopted both parts as ANSI/A3 R15.06-2025 in September 2025.
Credit
ISO technical committee 299, working group 3 (convenor Roberta Nelson Shea, Universal Robots), with CEN/TC 310; A3
Watch for
A standard for self-balancing robots
ISO working group 12 is drafting ISO 25785-1, safety requirements for industrial mobile robots that must actively keep their balance, such as legged humanoid robots, which "could become unstable in the absence of power." The European standards body CEN opened its matching project on 9 January 2026.
Achieved whenISO 25785-1 is published as an international standard.
Claimed by
ISO technical committee 299 and CEN/TC 310, through DIN, Germany's standards body, January 2026
Open
01
Measure a hobby arm's bump
A public table of the force and pressure that small, low-cost robot arms sold to hobbyists and schools put on a padded load cell when they bump into it at the speeds their makers allow, set against the body-region limits in ISO/TS 15066. Candidates are the SO-101 arm from Hugging Face's LeRobot project (see the Note "No download"), Elephant Robotics' myCobot 280 and UFactory's Lite 6. The test touches only the sensor, never a person, and certifies nothing.
The briefCredit · unclaimed02
AI risk checks, in hindsight
A test of whether an AI model, given only a description of a robot cell as it stood before an accident, lists the hazard that later hurt someone. Cases come from published accident accounts: NIOSH's 1984 die-cast case and other NIOSH investigations, and the free-text narratives in OSHA's public severe injury reports that mention robots. The outcome is removed before the model sees the case.
The briefCredit · unclaimed
Open 01 · the brief
Measure a hobby arm's bump
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table of the force and pressure that small, low-cost robot arms sold to hobbyists and schools put on a padded load cell when they bump into it at the speeds their makers allow, set against the body-region limits in ISO/TS 15066. Candidates are the SO-101 arm from Hugging Face's LeRobot project (see the Note "No download"), Elephant Robotics' myCobot 280 and UFactory's Lite 6. The test touches only the sensor, never a person, and certifies nothing.
Why it is still open
Published collision-force maps cover industrial cobots such as the UR10e and KUKA LBR iiwa, and a search found no published measurements for hobby or classroom arms, which are now used at home and in schools to record training data for AI robot models.
Built with what exists
A load cell with an HX711 amplifier board or a handheld force gauge; foam and rubber pads of known stiffness to stand in for the hand and forearm; a high-speed phone camera; LeRobot software to command repeatable moves.
First step
Build the load-cell rig and record 50 bumps from one SO-101 at its default speed.
Done looks like
For each of three arms, three speeds, five contact points and 10 bumps, the peak force in newtons and pressure in newtons per square centimetre, published with raw readings and the share of bumps above the ISO/TS 15066 hand and forearm limits.
Open 02 · the brief
AI risk checks, in hindsight
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test of whether an AI model, given only a description of a robot cell as it stood before an accident, lists the hazard that later hurt someone. Cases come from published accident accounts: NIOSH's 1984 die-cast case and other NIOSH investigations, and the free-text narratives in OSHA's public severe injury reports that mention robots. The outcome is removed before the model sees the case.
Why it is still open
Work on AI models for machinery risk assessment exists, including Padma Iyenghar's dataset of hazard scenarios from ISO 12100, the general machinery risk standard, with benchmarks published from 2023 to 2025, and SAFARI, a September 2026 benchmark of 3,000 automotive cases. F. Costantino and N. Sabetta used AI models in 2025 to sort OSHA accident narratives involving robots into recurring topics, such as maintenance and the body parts injured, after the fact. None of it tests whether a model can name the hazard from a description of the cell before the accident.
Built with what exists
OSHA's severe injury report data; the NIOSH alert; an AI model such as Claude Fable 5.2 (see the Note "Three working devices"); a second model that flags doubtful items for a person (see the Note "Seven more flagged"); a spreadsheet for scoring.
First step
Rewrite the 1984 NIOSH case and nine OSHA robot narratives as pre-accident descriptions and run one model on them.
Done looks like
60 cases, each rewritten as a pre-accident description by one person and checked by another, with published figures for the share of cases where the model names the mechanism that caused the injury, its false alarms per case, and the same two figures for two safety engineers working blind.