Table of Contents
Industrial AI Safety Standards for Machines That Think
Industrial AI safety standards were built on cages and hardwired stops. Humanoid robots that decide to sit down break that model. Here is what changes.
For sixty years, industrial AI safety standards did not need to exist, because industrial robots were kept in cages. The safety argument was physical and absolute: a fence, a gate switch, and a circuit that cut motor power the instant the gate opened. No perception, no inference, no learned model. Then on September 15, 2026, IEEE Spectrum published a piece headlined “Digit 5 May Be the First Humanoid Robot Worker That’s Truly Safe,” describing a 129-kilogram machine whose safety strategy is to notice people and sit down. Read that sentence twice. The safety function now depends on the robot being right about what it sees, and the entire certification framework was built on the assumption that it never would.
Key Takeaways
- Classical machine safeguarding is deterministic: OSHA’s Technical Manual describes fixed barriers, interlocked guards and presence-sensing devices whose job is to remove power, not to make a judgment.
- Agility Robotics’ Digit 5, unveiled September 15, 2026, uses a behavioral safety approach: the robot puts down its load and sits when a person approaches, which IEEE Spectrum’s Evan Ackerman framed as making it “physically impossible for Digit to fall over on anyone, ever.”
- OSHA’s robot guidance already recognizes graded approaches including speed and separation monitoring, power and force limiting, and hand-guided control, and caps teach-mode speed at 250 millimetres per second.
- NIOSH documented 41 robot-related fatalities in the United States between 1992 and 2017, with struck-by, caught-between, crushing, slips and electrical contact as the main mechanisms.
- OSHA’s updated warehousing National Emphasis Program, with inspections starting July 31, 2026, explicitly names robotics among the hazard categories inspectors will examine.
How industrial machine safety actually worked before this
A safeguard in a factory is not advice. It is a circuit.
OSHA’s Technical Manual on industrial robot systems lays out the hierarchy. Fixed barriers and perimeter guards physically prevent a worker from reaching the hazard. Interlocked barrier guards stop operation when opened, which means a switch on the gate is wired into the safety circuit so that opening it removes power. Presence-sensing devices such as light curtains, safety mats, scanners and vision systems detect an intrusion and stop the machine.
The crucial property is that these functions are designed to fail safe. A light curtain that loses power stops the machine. A broken wire in a gate interlock stops the machine. Safety engineers specify redundant channels and cross-monitoring so that a single component failure cannot leave the machine running with a person inside. That is the whole discipline, and it is why industrial safety has a good record despite the machines being genuinely lethal.
The manual also documents the hazards being guarded against: impact and collision from unpredicted movement, crushing and trapping of limbs between a robot and other equipment, struck-by from parts released by an end-effector, plus electrical, hydraulic and pneumatic hazards. And it specifies something very concrete for programming work. During manual mode teaching, “robot speeds should be placed at a reduced speed of 10 inches per second (250 mm/second) or less on any part of the application during teaching.” A number, in a government document, limiting how fast a machine may move while a human stands within its reach.
For maintenance, the controlling rules are OSHA’s hazardous energy standards, 29 CFR 1910.147 for lockout/tagout and 29 CFR 1910.333 for work practices. The manual also references ANSI/RIA R15.06 for industrial robot safety, ISO 10218-2 for power and force limiting requirements, ISO/TS 15066 for collaborative operation, and NFPA 79 for electrical standards on industrial machinery.
The four ways a robot and a person can share space
Collaborative operation is not one thing. It is a set of defined modes, each with its own verification burden. This table summarizes how they differ and what each one demands.
| Mode | What it means | What the safety system must prove | Where it is used |
|---|---|---|---|
| Safety-rated monitored stop | Robot halts whenever a person is in the shared space, then resumes | That the stop is reliable and that motion cannot restart while occupied | Machine tending, part loading |
| Hand guiding | The operator physically moves the robot using a hold-to-run device | That releasing the control stops motion immediately | Teaching paths, heavy part positioning |
| Speed and separation monitoring | Robot slows or stops as a person gets closer, based on measured distance | That the sensing covers the full approach and the stopping distance is bounded | Mobile robots, large cells |
| Power and force limiting | Contact is allowed but forces are capped below injury thresholds | That contact force stays under limits for every body region and geometry | Small assembly cobots |
Power and force limiting is the mode most people picture when they imagine a friendly robot, and it is also the most constrained. OSHA’s description is blunt: robots in this mode operate at “much lower speeds and payloads.” Physics does not negotiate. A machine that may legally touch a human has to be weak or slow or both.
Now place Digit 5 against that table. It is 1.8 metres tall, 129 kilograms, and carries 23 kilograms at a height of 2.1 metres. There is no power-and-force-limiting argument available for a machine like that. Agility’s answer, as reported to IEEE Spectrum, is different in kind: “As a person approaches Digit, Digit will put down whatever it’s carrying and then if necessary make sure that it’s stably seated on the ground before that person gets near.”
That is a genuinely clever safety concept. It is also a safety function whose input is perception. And the certification question nobody has fully answered is how you demonstrate, to the standard that industrial safety normally requires, that a perception-plus-decision pipeline will behave correctly in a situation no one tested.
Why a learned model is hard to certify, in plain terms
Functional safety assessment works by enumerating failure modes and bounding them. You ask: what can break, how likely is each break, and does the system still stop the machine. For a relay you have failure rate data. For a light curtain you have a tested response time. For a redundant dual-channel circuit you can compute a probability of dangerous failure per hour.
A neural network does not decompose that way. Its failure mode is not “channel A opened.” Its failure mode is “the input was outside the distribution it was trained on and the output was confidently wrong.” You cannot enumerate that set, because it is defined by everything the training data did not contain. A worker in an unusual high-visibility jacket under unusual lighting carrying an unusual object is not a component failure. It is a gap in a dataset.
The engineering responses available today are all about containment rather than trust. Keep the learned component out of the safety function and let it handle productivity only. Put a conventional, deterministic layer underneath it, such as a scanner-based protective stop that does not care what the network thinks. Bound the physics so the worst case is survivable regardless of the decision, which is exactly what Agility is doing by ensuring the robot cannot topple onto someone. Reduce the state space by designing the task so fewer novel situations can arise.
Notice that three of those four are the old discipline applied to a new machine. Which is the honest read on where this stands: nobody has figured out how to certify machine learning as a safety function, so the field is instead building machines whose worst case is tolerable. That is good engineering and it is not the same thing as a solved standards problem.
Boston Dynamics’ October 1, 2026 announcement of a new Atlas hand fits this pattern. IEEE Spectrum described it as a specialized gripper that may outperform more humanlike designs, balancing capability against reliability and manufacturability. Reliability is a safety property. Choosing fewer, sturdier fingers over more dexterous ones is a safety decision dressed as a design decision.
What this means for workers and for inspections
Regulators are moving, and the direction is enforcement rather than new permissions.
OSHA’s warehousing and distribution centre page names the hazard categories for its National Emphasis Program: powered industrial trucks, ergonomics, material handling, hazardous chemicals, slips, trips and falls, and robotics. Inspections under the updated program began July 31, 2026. Robotics is on a list with forklifts, which is a plain statement that inspectors will be looking at robot installations the way they look at any other machine hazard.
The background injury data is better than people assume. BLS reported on January 22, 2026 that private industry employers recorded 2.5 million nonfatal workplace injuries and illnesses in 2024, a rate of 2.3 cases per 100 full-time equivalent workers, down from 2.4 and the lowest in the series going back to 2003. Manufacturing was among the sectors whose rate decreased. Industrial safety practice works. The question is whether it transfers to machines that make decisions.
The historical robot-specific record comes from NIOSH, which documented 41 robot-related fatalities in the United States between 1992 and 2017 and lists the mechanisms as struck-by or caught-between, crushing and trapping, slips, trips and falls, and electrical hazards. NIOSH also flags emerging risks that are not mechanical at all: unexpected contact injuries, worker distraction, and psychological stress related to job security. That last item is unusual for a safety agency to name and it deserves more attention than it gets.
Against that backdrop, the growth figures matter. NIOSH reports industrial robots in U.S. factories grew 10 percent in 2022, and professional service robots sold in the United States reached 158,000 units in 2022, a 48 percent increase. More machines, in less structured settings, near more people.
What to Watch For Over the Next 3 Months
- Week 4: Watch for any published safety case for a humanoid in an industrial setting. Not a blog post, a safety case: hazard list, risk assessment, the specific protective measures, and the residual risk. If that document starts appearing, the field is maturing. If it does not, deployments are running on vendor assurance.
- Month 2 red flags: A deployment where the only protective measure is the robot’s own perception, with no independent layer. Independence is the oldest principle in this field, and a system that monitors itself and reports that it is fine has no independence at all.
- Month 3 self-check: Ask whether you can explain the difference between a robot that stops because a circuit opened and a robot that stops because it decided to. If you can, you understand why this is a standards problem and not a software problem.
Frequently Asked Questions
Are collaborative robots actually safe to work next to?
In the defined modes, yes, with a caveat: safety belongs to the application, not the robot. A power-and-force-limited arm holding a sharp tool is not safe merely because the arm is rated. OSHA’s framework requires a task-based risk assessment of the whole installation, including the tool, the part and the fixture.
Has anyone been killed by a collaborative robot?
NIOSH’s documented count of 41 robot-related fatalities in the United States from 1992 to 2017 predates widespread collaborative deployment and is dominated by conventional industrial robots, typically during maintenance or setup when guarding was bypassed. That pattern, incidentally, is the strongest argument for lockout/tagout discipline.
Who writes these standards?
A mix. ISO publishes the international robot safety standards including the 10218 series and the ISO/TS 15066 technical specification for collaborative operation. In the United States the national versions have historically carried ANSI/RIA designations, now under the Association for Advancing Automation. OSHA enforces its own general duty and hazardous energy rules and points employers to those consensus standards.
Should a parent worry about their teenager working near robots?
Worry less about the robot and more about the training. The dangerous moments in the historical record are maintenance, setup and troubleshooting, when someone is inside the hazard zone with the guards off. The question to ask an employer is who authorizes entry into a robot cell and how lockout is verified.
Does any of this apply to home robots?
Partly, and the gap is uncomfortable. Industrial standards assume a trained worker in a controlled environment. A household has children, pets, stairs and clutter, and none of the mode definitions above were written for that. Our look at home humanoid robots covers what is actually shipping, and our piece on cobot programming careers covers who does this work.
About the author
Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.
Sources
- Occupational Safety and Health Administration. “Industrial Robot Systems and Industrial Robot System Safety.” OSHA Technical Manual, Section 4, Chapter 4. U.S. Department of Labor. https://www.osha.gov/otm/section-4-safety-hazards/chapter-4
- Ackerman, E. (2026). “Digit 5 May Be the First Humanoid Robot Worker That’s Truly Safe.” IEEE Spectrum, September 15, 2026. https://spectrum.ieee.org/humanoid-robot-safety
- National Institute for Occupational Safety and Health. “Center for Occupational Robotics Research.” Centers for Disease Control and Prevention. https://www.cdc.gov/niosh/robotics/about/index.html
- Occupational Safety and Health Administration. “Warehousing and Distribution Centers.” U.S. Department of Labor. https://www.osha.gov/warehousing
- U.S. Bureau of Labor Statistics. (2026). “Employer-Reported Workplace Injuries and Illnesses, 2024.” News release, January 22, 2026. https://www.bls.gov/news.release/osh.nr0.htm
- The Robot Report. (2026). “Top 10 robotics stories of September 2026.” October 1, 2026. https://www.therobotreport.com/top-10-robotics-stories-of-september-2026/
- Ackerman, E. (2026). “Video Friday: Two Birotors Make a Quadrotor.” IEEE Spectrum, September 18, 2026. https://spectrum.ieee.org/video-friday-quadrotor-from-birotor
