Friday, November 20, 2026

When a Robot Hit Me on the Head: The Story of an Experiment

From the early history of the Area of Normal Functioning (ANF)

About a year ago, I visited Japan and went to the Toyota Museum. Among its many fascinating exhibits was a display of industrial robots representing several generations of robotics, from the 1980s to the present day.

The latest robots were actually working, demonstrating automobile assembly operations right in front of visitors.

But the first robot in that historical lineup was an old Kawasaki Unimate.

When I saw it, I became unexpectedly emotional.

It was almost like meeting an old friend!

Well, not exactly the same robot, but one very similar to a machine I had worked with almost forty years earlier while conducting robotics research at the Leningrad Polytechnic Institute, now Peter the Great St. Petersburg Polytechnic University.

The robot I worked with was a Soviet industrial robot called Universal-15M, a large hydraulic machine installed in the Department of Mechanical Engineering Technology, next door to my own department.

And I have a rather amusing story about that robot.

Although, at the time, it wasn't funny at all.

Why I Needed a Robot

In the late 1980s, I was conducting research on mathematical modeling of robotic systems.

One of the important concepts I used in that work was the Area of Normal Functioning (ANF).

The basic idea was that the successful completion of a technological operation depends not only on correct control commands but also on the initial state of the system, geometrical deviations, mechanical characteristics, and interactions among its components.

By building a mathematical model, we can determine a region of initial conditions within which an operation will be completed successfully.

In the Soviet Union, theoretical results alone were not sufficient for a dissertation. Experimental validation was a mandatory requirement.

So, besides deriving equations and writing computer programs, I also had to build experimental equipment, perform measurements, and work with real industrial robots.

This experiment was part of my dissertation research. The photograph of the experimental setup comes directly from my dissertation, a copy of which I still have today.

Actually, I used several experimental setups during my research, but this particular one is the subject of today's story.

The Experiment with a Lathe

The task was relatively simple.

The robot had to pick up a cylindrical metal workpiece — a shaft — and insert it into the chuck of a lathe.

The chuck had three jaws, initially in the open position. The robot brought the workpiece into position, the jaws closed around it, and then the robot released the shaft so machining could begin.

Sounds straightforward, doesn't it?

But imagine that the robot positions the shaft slightly too high, too low, or sideways relative to the correct position.

What happens when the chuck jaws begin to close?

In some cases, the jaws can correctly position and secure the workpiece. In others, the shaft may become misaligned, jam, or end up clamped with an unacceptable positioning error.

This was exactly the problem I was investigating.

First, I developed a mathematical model describing the mechanical interaction between the robot gripper and the lathe chuck.

The model included a system of nonlinear differential equations describing motion, contact forces, and the different phases of the gripping and clamping process.

I wrote Fortran programs to solve these equations numerically.

Using what I called the trajectory method, I calculated the region of permissible initial workpiece positions for which the loading and clamping operation would end successfully.

This was the Area of Normal Functioning for that particular operation.

But the calculations needed experimental verification.

Sensors, Measurements, and Experimental Validation

For this purpose, I designed and built a special experimental fixture, visible in the old photograph from my dissertation.

Experimental setup for studying the Area of Normal Functioning of the robot gripper–lathe chuck system. Late 1980s. Photograph from my dissertation

I needed to vary the initial position of the shaft relative to the chuck along two coordinates, measure those deviations, and then evaluate the result of the clamping operation.

The fixture included mechanical dial indicators for measuring displacement.

One of these indicators measured the vertical coordinate. It was mounted underneath the assembly and can be seen in the lower-right corner of the photograph.

The experimental procedure was straightforward but rather labor-intensive.

I would position the shaft at a selected initial location, measure its coordinates, activate the clamping mechanism, and observe the result.

Was the shaft properly secured? Did jamming occur? What was the final positioning error?

I repeated these experiments many times, changing the initial conditions.

The experimental results were then compared with the ANF boundaries calculated using my computer model.

In my 1990 paper, Computer Analysis of the Quality of Functioning of Piece-Transfer Systems, I presented the mathematical model and the results of its experimental validation.

The differences between calculated and experimental results did not exceed 20%.

For a fairly complicated nonlinear mechanical system, that was an encouraging outcome.

And Now Comes the Interesting Part

The Universal-15M was a hydraulic robot. As far as I remember, the Kawasaki Unimate used a similar type of drive system.

Hydraulic systems could generate considerable forces, but they required reliable components and proper maintenance.

During my experiments, I periodically had to adjust the measuring instruments.

The most inconvenient instrument was the lower dial indicator used to measure vertical displacement.

To adjust it, I had to put my head directly underneath the robot's arm.

And one day, while I was making this adjustment, something completely unexpected happened.

The robot suddenly lowered its arm and hit me on the head!

It didn't drop its arm completely, of course. But it moved downward far enough to give me a rather painful blow.

Apparently, something had gone wrong in the hydraulic system.

It was a complete surprise. One moment everything appeared to be operating normally, and the next moment — bang!

It hurt quite a bit.

Fortunately, I wasn't seriously injured.

What I remember most clearly is that I felt more offended than frightened.

Imagine that!

Here I was, doing serious scientific research, developing mathematical models, solving nonlinear equations, and studying the proper functioning of this robotic system.

And the robot had the nerve to hit me on the head!

Clearly, it had no respect for the scientist conducting the experiment.

Everything Is Normal… Until It Isn't

Almost forty years have passed since that experiment.

Today, I am once again working with the concept of the Area of Normal Functioning, but in a much broader context.

My current research extends beyond mechanical and robotic systems to computer infrastructure, artificial intelligence, organizational processes, and even human behavior.

Looking back at that incident, I now see a certain irony in it.

My mathematical model described in considerable detail the interaction between the cylindrical workpiece and the jaws of the lathe chuck.

It allowed me to calculate the initial conditions under which the loading operation would be completed successfully.

But the unexpected downward movement of the robot arm, apparently caused by a hydraulic malfunction, was not part of that model.

And, of course, the model certainly did not include the possibility of a researcher's head being positioned underneath the robot arm!

This is an important distinction.

An Area of Normal Functioning is always defined for a particular system, under particular conditions, and according to specific criteria of successful operation.

A system may function normally with respect to one task while simultaneously creating hazards associated with other factors that the model does not account for.

Therefore, a mathematically defined ANF does not necessarily coincide with the region of safe operation of the entire system.

This distinction has become increasingly important in my current work.

It is not enough to ask whether a system is behaving normally according to its model. We must also ask whether the model includes the relevant failure modes, external constraints, and safety requirements.

A system can be operating inside its modeled normal region while still being unsafe in a broader context.

And that brings me back to the title of my forthcoming book:

Everything Is Normal… Until It Isn't.

Sometimes we discover that a system has crossed the boundaries of normal functioning not from a dashboard, a mathematical model, or a warning signal.

Sometimes we discover it quite literally — with our own head.

No comments:

Post a Comment