Framework
What ‘avoids obstacles’ actually claims
Three words on a spec sheet stand in for four separate engineering problems: noticing something is there, knowing what it is, deciding what to do, and recovering when the decision is wrong. Most products solve the first and stumble on the last.
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.
Two pipelines, not one
Mapping space versus naming things
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 honest 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.
Reading the ladderWhy the gap changes the classification
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
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.
