← All research

Risk research

Pricing robot risk without a loss history

Insurers price risk from loss history, and robots don't have one yet. We tested how far public data and causal methods can close that gap, and what needs to be measured going forward.

Every insurance line starts from the same place: a long record of what went wrong, how often, and what it cost. Car insurers have a century of crashes to learn from. A company deploying its first fleet of warehouse robots or humanoids has a few months, if that.

This is the cold-start problem: a risk has to be priced before the history that would normally set its price exists. It is also a real constraint on adoption, because what can't be measured is hard to underwrite.

To understand how far public data can go, we assembled 3.1 million OSHA workplace injury records, California's autonomous vehicle filings, nearly 3,000 federal crash reports involving automated vehicles and 271 million miles of Waymo driving data. We then evaluated two established causal-inference methods, first on simulated fleets where the true answer was known, then on the real data.

Our main findings:

  • Naive comparisons can reverse the answer. Robots are deployed selectively, usually into easier work, so raw incident rates can point the wrong way.
  • Near misses are the strongest early signal, but only under a shared definition. Today's reporting makes them incomparable across operators.
  • A safety improvement can't be verified at a single site. Pooling sites that make the same change makes it measurable.
  • Some risk structure transfers across robots, so a new robot doesn't have to be priced from zero.

1. Naive comparisons can reverse the answer

Robots are not assigned work at random. Operators enable autonomy where conditions are calm and keep a person closer when they aren't. New robots go into new buildings. A raw comparison of incident rates largely measures these deployment choices, not the robot.

We first tested this in simulation. We built simulated fleets in which autonomous operation was, by construction, 45% riskier than supervised operation. A raw comparison of incident rates estimated autonomy to be 18% safer. Two methods that compare robots under matched conditions, Genetic Matching and regression adjustment, both recovered the true effect to within about 6%.

We then applied the same approach to real data. OSHA publishes serious-injury counts and hours worked for every large warehouse each year. In 2020, the investigative newsroom Reveal published a large online retailer's internal labels identifying which of its fulfillment centers use mobile warehouse robots. Together, these give 1,342 site-years across 214 buildings.

In the raw data, robotic sites appear 20% safer for workers. But most non-robotic sites handle heavy, bulky goods, the most injury-prone work in the network, while robots were deployed mainly in small-item sites. Once sites are matched on product type, region, size and age, the result reverses: robotic sites had up to 55% more serious injuries per hour worked. Three of the four like-for-like estimates sit clearly above parity, consistent with Reveal's reporting. The gap has narrowed over time: among small-item sites it fell from 8.4 versus 5.6 serious injuries per 200,000 hours in 2019 to 4.6 versus 4.2 in 2025.

Raw data says robots cut injuries. Like-for-like says the opposite. Serious-injury rate at robotic sites ÷ comparable non-robotic sites, with 95% intervals 0.6×1.0×1.4×1.8×2.2× Same rate Raw: all robotic vs all non-robotic 0.80× Same product type · matching 1.55× Same product type · regression 1.47× All labelled sites · matching 1.34× All labelled sites · regression 1.14×
Boop analysis of OSHA injury logs (2017–2025) joined with Reveal's 2020 Amazon facility labels.

The data do not identify the mechanism; one plausible explanation is work pace rather than physical contact. The implication for pricing holds either way: a comparison that doesn't hold the work constant can get the sign of the effect wrong, and here it did.

2. Near misses are the best early signal, if measured consistently

Near misses are the natural way around a short incident history: they occur far more often, so they accumulate into usable data much sooner. In our simulations, near misses estimated risk several times more precisely than incidents from the same fleets.

This only holds if an operator's near-miss rate predicts its collision rate. California is the only jurisdiction that publishes both for the same fleets: 18 autonomous vehicle operators, 17.6 million testing miles, about 86,000 disengagements (a safety driver taking over) and 518 collisions.

In this data, it did not. Operators with more disengagements per mile did not have more collisions per mile, and the ratio of collisions to disengagements varied roughly 260-fold across operators.

The main driver is definitional. Each company sets its own threshold for what counts as a disengagement. Apple reported 78,971 over the period we studied, against Waymo's 1,065, while driving about a ninth of Waymo's miles. This reflects reporting practice, not relative safety.

Near misses are the right early signal for robot risk, provided they are measured the same way on every robot.

3. Verifying a safety improvement takes more than one site

Operators reasonably want credit for safety improvements: a firmware release, a new guarding layout, a speed limit near people. Granting that credit fairly requires measuring the effect.

The standard method is synthetic control: build a counterfactual "twin" for the site that changed from a weighted blend of comparable sites that didn't, then compare outcomes. We applied it to a real policy change: California's warehouse quota law, AB 701, which took effect in January 2022 and restricts quotas that interfere with rest breaks.

After the law took effect, the serious-injury rate at that retailer's California sites fell about 17% more than its synthetic twin's, consistent with a real effect. But when we ran the same test on each of the 21 comparison states as placebos, 2 showed a drop as large. With a single treated state, the effect can't be distinguished from normal year-to-year variation.

Our simulations show this more sharply. At a single site, a true 30% reduction was often missed; pooling five sites that made the same change detected it almost every time. A simple before-and-after comparison performed worst, reporting changes at sites where nothing had changed.

The practical implication: safety improvements should be measured across a vendor's release at several customers, not at a single site. New York, Washington and Minnesota have since passed similar laws, which will allow a pooled version of this test as their data becomes available.

4. Some risk structure transfers across robots

Not everything has to be learned from scratch for each new robot. Three regularities carried over in the data.

How bad a crash is depends on what was hit. Across nearly 3,000 federal crash reports on driverless vehicles, 41% of crashes involving a pedestrian, cyclist or motorcyclist led to an alleged injury, against 10% for cars and 3% for objects. The vehicle's own speed had no measurable effect. Severity follows the situation, so it can be shared across robot types instead of re-learned for each.

Mature systems behave consistently across locations. Waymo's crash rate varies only about ±15–18% across the 8 counties it operates in. Territorial rating, the basis of human auto insurance, is a poor fit for robots.

Validated simulators transfer. In our simulations, pricing a new robot type from its specifications was off by a factor of nearly four. A simulator whose scores closely track real-world risk priced it as accurately as years of operating history.

A small fleet's own record accumulates slowly. In our simulations, a 20-robot fleet needed three to five years of incidents before its own data could price it with confidence. Until then, the most reliable price comes from what transfers: shared severity patterns, the broader portfolio and a validated simulator.

What this means for robot risk data

The methods held up. Applied carefully, public data reversed a conclusion the raw numbers would have given, and declined to confirm an effect that a simpler test would have reported. In every case, the binding constraint was what had been recorded in the first place.

Some of the most consequential quantities are not recorded publicly at all. In our model of a warehouse fleet, most of the expected cost of robot incidents comes from downtime, not worker injuries. Yet how often robots cause loss events on a warehouse floor isn't published anywhere, and the price varies 12-fold depending on that single number.

The study points to four things a robot's data must capture before its risk can be priced fairly:

  • Exposure, not just incidents: hours, task, payload, speed, people nearby, autonomy mode and software version, so comparisons can hold the work constant.
  • Near misses as first-class events, under one definition across every robot.
  • Enough history before and after a change that a safety upgrade can be measured across several sites and credited fairly.
  • For every incident: who or what was hit, both parties' motion, contact force and downtime, because that's what drives the cost.

Boop's Robot Data Recorder is designed to capture what isn't recorded today: a single near-miss standard and a structured record of the circumstances around every incident.

Our next study tests simulators against robots that already have a field record. If simulator scores track real incident rates closely enough, a new robot can be priced fairly from day one rather than after years of accumulated history.

Limitations

  • Simulations test methods, not the world. Effect sizes within them, such as autonomy being 45% riskier, are parameters we set to check whether each method recovers a known answer.
  • The warehouse result covers one company's robots of one generation, with 10–51 comparison sites. It shows an association, not proof that the robots caused the injuries.
  • The crash data covers one kind of robot, self-driving cars, and California's figures come from testing with safety drivers.
  • Matching only corrects for what's recorded. Anything left unmeasured can still bias a comparison.
  • These results have not yet been peer-reviewed. We are sharing them to invite scrutiny of the methods and data.

Sources

If you deploy robots and want to collaborate on this work, we'd like to hear from you.

Get in touch

More research