Skip to content Skip to footer

The robot makes fewer mistakes. Who takes the hard cases?

In brief. A robot can improve by avoiding the riskiest moves. To understand how much it pays, you also have to count the work it leaves to operators. In the experiment published by NTT on November 11, 2026 the success rate rose from 73.1% to 92.3%.

If the robot leaves all the difficult parcels on the belt, its success rate can go up. The shift can become harder for whoever has to deal with those parcels. When I read a success percentage, I also want to know what ended up outside the count.

This is a question to bring to an acceptance test, not a description of NTT’s result. In the experiment announced on November 11, 2026, the Japanese company had a small robot stack objects on a shelf. With a model that anticipates the outcomes of movements, the success rate went from 73.1% to 92.3%. The first intended use concerns airport baggage, with field trials to be developed together with Mujin. The published result still concerns the shelf. NTT · Experiment and results

The interesting step is that the robot assesses the consequences before moving. Among several possible gestures it looks for the one most likely to succeed, also considering what can happen to nearby objects. Putting a box down is not just getting it to the right spot: it has to stay there, without knocking the others over.

A good grip, seen alone, can be a terrible decision inside a stack. It is this relationship between actions that makes the test interesting even for those who do not work in robotics.

The box next to it

In the Scenari conversation with Simone D’Aquino of Capgemini we discussed how to bring physical constraints into design models. The image on the screen has to reckon with the behavior of the object. The same problem returns when a machine has to act: the shape it sees does not automatically tell everything that will happen after contact. Scenari · Augmented engineering with Capgemini (in Italian)

In the digital world we can repeat many tests changing one parameter. In the physical world an error can leave something on the floor, interrupt a flow and require a person to step in. The benefit of anticipating it therefore includes what will not need to be recovered.

The two values in the NTT test are 19.2 percentage points apart. It is an evident improvement within that test. To turn it into an industrial calculation, though, other elements are missing: the time per operation, the variety of parcels, the frequency of interruptions and the work required when something fails. The percentage opens an assessment; it does not close it.

A robot that takes longer to choose can pay off if it avoids a long stoppage. It can pay off less when the flow is regular and the error is easily recovered. The cost of the gesture depends on what happens around it, even if the vendor manages to measure it with great precision.

Leaving the hard part to people

Whoever buys automation may want exactly this: take the repetitive part away from people and entrust to their experience what the machine cannot yet handle. Leaving them the exceptions can be the right design. There is no reason to demand that every machine be able to do everything before putting it to work.

I agree. But then the role left to people has to be designed, not discovered during the shift. If they used to alternate easy and difficult picks and afterwards receive only the difficult ones, the load is no longer measured by the total number of pieces. Fatigue, the attention required, skills and recovery times all change.

The benefit can still be there, even a large one. It has to include that transformation. An investment that frees ordinary hours but concentrates all the exceptions on a single operator creates a new dependency. It only takes that person being unavailable for part of the advantage to disappear.

To the acceptance test, therefore, I would bring the packaging that really arrives at the plant: the deformed ones, the soft ones, the combinations that make a stack unstable. The sample has to resemble the work, not a catalogue photograph. And next to the machine you have to observe whoever steps in, with the space and tools they will have in the real situation.

The count has to follow the parcel until it returns to the flow. If the robot rejects it and a person resolves it ten minutes later, that time belongs to the process. If instead a slower choice avoids an interruption for everyone, the advantage has to be recognized even if the single pick looks less fast.

Recovering from an error also produces information. The operator may recognize a package that often breaks or a combination the system handles badly. If that detail stays only in their experience, the machine will keep presenting the same problem. The vendor has to explain how these cases enter the updates and how you verify that the fix worked.

A test over a complete shift makes visible build-ups that a short demo can hide. The parcels set aside do not disappear when the video ends. Someone has to work them, often when there is less time left.

NTT’s robot shows a promising way to reduce errors before the action. The investment decision comes later, when we know how that improvement changes the day of whoever works beside it. The result that matters to the company is the work completed by the whole shift, including the boxes the machine chose not to touch.

Read next The test comes before the code

Background Impact · the case file (in Italian)

Image: NTT