The hidden floor: what safety actually means for consumer robots

Cliff sensors are standard on robot vacuums. The implementation uses infrared emitters and receivers under the chassis. When the IR signal fails to reflect within an expected range, the robot reverses. That logic is reliable on mid-toned, matte flooring. It is not universal.

Dark or absorbent surfaces, including black hardwood, deep navy carpet, and certain matte finishes, can absorb enough IR that the sensor misreads a valid surface as a drop. The opposite failure also exists: polished tile and glass-topped stair landings return ambiguous signals. Sensor contamination from dust and hair on the IR window is a third failure mode, often invisible to the owner until an incident occurs.

Edge geometry introduces a further variable. Stair edges with rounded bullnose profiles return ambiguous IR depending on approach angle. Incidents cluster around beveled nosing rather than square edges precisely because the calibration window was designed for square geometry. Premium models add redundant detection layers: multiple IR arrays, software confirmation, and downward-facing cameras that cross-check the primary reading. Budget models typically run a single IR pair per side.

iRobot introduced AI-based pet waste avoidance on the j7+, backing it with the Pet Owner Official Promise, an explicit replacement guarantee tied to the feature. Roborock, Dreame, and Eufy followed with their own implementations. Marketing language for all of them describes detection as robust avoidance, which is accurate under nominal conditions and incomplete everywhere else.

The underlying technology is a trained classifier. It carries a true-positive rate and a false-negative rate. Owner reports document failures in low light, with unusual debris geometry, with debris partially under furniture, and during edge-case combinations the training data did not anticipate. None of this is surprising. It is how classifiers behave when the input distribution shifts.

If your Roomba j7+ robot doesn’t avoid pet waste, we’ll replace it.

iRobot, Pet Owner Official Promise

That commitment converts a probabilistic capability into a verifiable contractual obligation with a defined consequence. Most other manufacturers describe the feature without equivalent guarantees or stated performance conditions. The practical implication for buyers is to treat pet detection as a useful but imperfect layer. Supervising a robot vacuum’s first runs in a household with pets remains a sound baseline regardless of which model is running.

Lawn robots present a categorically different safety surface because the cutting mechanism causes serious injury. The architecture has converged on several layers: lift sensors that cut blade power on tilt, bumper-strike behavior that triggers reversal on contact, virtual boundary obedience via RTK or buried wire, and remote disable from the manufacturer app. Premium RTK mowers add obstacle and pet detection as a further margin.

Robotic lawn mowers present a distinct hazard profile from traditional push mowers because they operate unattended and may resume operation after a brief pause without operator awareness.

CPSC, Consumer Product Safety Review, 2023

Despite multiple protective layers, lawn mowers remain the most dangerous consumer robot category. The blades are sharp, the operating pattern is unsupervised, and a household with small children or free-roaming pets accepts non-zero risk regardless of feature support. Manufacturer guidance consistently recommends direct supervision in mixed-use yards, which is the honest disclosure that safety architecture alone cannot substitute for oversight.

Pool robots operate in conductive water, so the safety surface is dominated by electrical isolation. Reputable pool robots run on low-voltage DC supplied through a transformer at the deck. The robot itself does not carry household-voltage current, and GFCI circuit protection is standard wiring code in most U.S. jurisdictions. A correctly wired installation means a fault in the robot trips the breaker rather than energizing the water.

Corded models can tangle on pool features including ladders and drain covers, which owner reports cite as the most common operational frustration. Cordless models eliminate that risk but add retrieval and charging steps. Salt and aggressive chlorine chemistry accelerate seal degradation in models not rated for those conditions. Pool robot safety is the most mature story in the consumer category because the failure modes are well understood and the residual risks are primarily owner-side.

The Robovations safety framework applies consistently across all four categories. Documented incidents from owner reports and recall notices form the primary evidence base. A robot with a recall history requires a longer recovery period before regaining its prior safety classification. Manufacturer disclosure practices are evaluated separately. Manufacturers that publish failure rates, edge cases, and explicit guarantees score higher than those issuing blanket safety claims without conditions.

Redundancy is the structural signal. Robots with multiple independent detection layers score higher than those depending on a single sensor type. Recovery behavior matters equally: a robot that stops, alerts, and waits when a sensor fails is classified differently from one that continues operation through the error. The framework is implementation-agnostic. Older sensor technology with sound engineering can score well. State-of-the-art AI with undocumented failure modes can score poorly.

Safety claims deserve the same scrutiny as autonomy claims. Avoids stairs, avoids pets, safe for children: each phrase is a marketing simplification of a probabilistic engineering claim. The relevant questions are under what conditions the claim holds, at what failure rate, and with what recovery behavior when the stated condition is not met. Those questions are rarely answered in the product listing.

The floor varies by category and household. A small home with no pets and no children tolerates a wider range of safety profiles than a home with both. The same robot may be appropriately safe in one setting and a genuine liability in another. The classification framework cannot make that determination for the buyer. It can make the underlying evidence visible enough that the decision can be made deliberately, with the failure modes on the table rather than buried in a footnote.