Framework
Do robot vacuums really avoid obstacles? What the claim actually means
Partly. The claim usually covers only detection, the cheapest of four steps: detection, classification, avoidance, and recovery. Most robots sense that something is ahead but misjudge what it is and fail to recover when wrong, and recovery, the step that decides usability, rarely appears on the box.
Almost every consumer robot sold today carries some version of the phrase. Avoids obstacles. Recognizes obstacles. AI obstacle avoidance. It reads as a single capability, a box that is either checked or not. It is not one capability. It is a sequence of four, and a robot can be excellent at the early steps while failing the ones that decide whether your charging cable survives the week.
Separating the steps is the whole point of classification. A spec line collapses them into a marketing claim. The ladder below is how we read that claim back apart.
Figure
What obstacle avoidance actually covers
- The thresholdThe smallest height the sensor resolves. Everything below this line is driven over rather than avoided, whatever the obstacle is.
- Avoidedtall enough to register before contact
- Driven intobelow the sensor floor, whatever the marketing says
Anatomy of a claimThe four steps hidden in one phrase
Detection is the floor. The robot registers that something occupies space ahead, through a bumper, an infrared cliff sensor, ultrasonic ranging, structured light, or LiDAR returns. Detection answers one question: is there something there? Nearly every robot above the budget tier clears this step. It is also the cheapest to advertise, which is why the weakest products lean on it hardest.
Classification is where products separate. Detecting a mass at coordinates is not the same as knowing it is a sock, a cord, a pet, or a wall. Classification needs a camera feed and a trained model running against it.
Roborock draws this line explicitly in its own product language: a reactive system without a neural network detects an obstacle and runs a fixed response, while the version with a convolutional model attempts to identify the object and react differently to a shoe than to a pet. The difference is not marketing polish. It is whether the robot can choose a response that fits the thing in front of it.
How to read the claim
Four steps, four different promises
| Step | What it answers | What it needs | Where products fail |
|---|---|---|---|
| Detection | Is something there? | Bumper, IR, ultrasonic, LiDAR | Rarely fails; cheapest to claim |
| Classification | What is it? | Camera plus trained model | Many robots stop here |
| Avoidance | What do I do about it? | Path replanning around the labelled object | Replans into a new dead end |
| Recovery | What if I was wrong? | Self-extraction, error reporting, learning | The step almost nobody advertises |
Avoidance is the action. Once an object is detected and ideally classified, the navigation stack has to replan a path around it without abandoning the job. This is harder than it sounds. A robot that swerves around a cable into a chair leg has avoided one obstacle by creating a second problem. Good avoidance is measured by what happens to the rest of the run, not by the single dodge.
Recovery is the step that almost never appears on a box. Every classifier misfires. The question is not whether a robot avoids obstacles but what it does in the seconds after it guesses wrong: does it wedge itself under a couch and run the battery flat, or does it detect that it is stuck, reverse, and report the failure clearly enough that you can fix the cause?
Recovery is where a $400 robot and a $1,400 robot diverge most, and it is the step a spec sheet is least likely to mention.
[rv_ans_score_pillars robot=”roborock-s8-maxv-ultra”]
Reading the ladderDetection and recovery are different classifications
On the Autonomy Ladder, the difference between detection and recovery is the difference between a robot that needs supervision and one that can be left alone.
A Level II robot can detect and sometimes avoid, but it depends on a tidy floor and a forgiving room. A robot earns a higher classification not by detecting more obstacles but by handling the consequences of its own mistakes without a human in the loop.
This is why we treat the four steps as separate evidence. A manufacturer claim about recognition tells us a model exists. It does not tell us the recovery behavior, and recovery is the part owners actually live with.
Steps a spec line usually proves
1 of 4
Detection is easy to claim and easy to verify. Classification, avoidance, and recovery each need separate evidence the box rarely provides.
Detection in isolation
A cliff sensor only answers one question
- Floor returnsThe pulse leaves, meets a surface within range, and comes back. The sensor reads floor and the robot drives on.
- Dark floor scattersA matte black surface absorbs the pulse instead of returning it, so nothing comes back.
- A real dropNothing within range to return from. The signal is identical to the rug, which is why a robot can refuse to cross a dark mat.
The practical takeaway is a habit, not a purchase. When a product says it avoids obstacles, ask which of the four steps the evidence actually covers. The phrase is doing four jobs. Most of the time it has only earned the first.
A robot that detects everything and recovers from nothing is still a robot you cannot leave alone. Recovery is the step that decides the classification.
