The human hand has 35 actuators  or muscles for producing movement.

19 of which are inside the hand - responsible  

for fine movements like bringing the  fingers together or pulling them apart.

The other 16 are located outside the hand,  

but extend into it via tendons. These  provide power for gripping and lifting.

Scientists can reproduce at least some of the  hand’s physical elements and capabilities.

But the rub has been control. How  do we convey to the prosthetic all  

the complicated movements behind  a "simple" gesture like grasping?

In the early 1960s, a Yugoslav scientist  from the well-known Mihajlo Pupin Institute  

named Rajko Tomović looked  to biology for a solution.

He leveraged his findings to create arguably the  

first autonomous robotic hand. Certainly  the first such five-fingered robotic hand.  

And it was capable of precision  tasks that might surprise you.

In today’s video, we profile a series  of pioneering robotic prosthetics from  

Communist Yugoslavia: The Belgrade Hands.

## Beginnings

Tomović (Рајко Томовић) was born  in November 1919 in Baja, Hungary.

In 1936, his family settled in Belgrade  where he did his secondary schooling.  

He then enrolled at the University of  Belgrade to study electrical engineering,  

but those studies were  interrupted by World War II.

Tomović joined the Yugoslav Communist Party  and fought the Axis forces. He was arrested,  

sent to a forced labor camp, and for some  time worked in a mine. At the end of the war,  

he had risen to be a captain  in the Yugoslav People's Army.

After the end of the war, he transferred  back to the University of Belgrade to  

finish his studies. He earned his doctorate  in the field of analog computers and then  

worked for a decade in Yugoslavia's  nuclear research institute in Vinca.

In 1960, Tomović left - perhaps due to Tito  halting the Yugoslav nuclear weapons program.  

He worked at UCLA in 1961 as a visiting scholar  and then joined the Mihajlo Pupin Institute.

There he helped produce one of the  country's first computers - the vacuum  

tube and transistor-based CER-10. After that,  he pivoted into work for developing prosthetics,  

reportedly due to a desire to help  injured World War II veterans.

## Controlling the Hand

People have been making hand prosthetics  since at least the days of the Roman Republic.

In 218 BC, the Roman general Marcus Sergius  is said to have been fitted with an iron  

prosthetic that allowed him to hold a shield  and fight battles against Hannibal of Carthage.

In the 1500s, two iron prosthetics  were built by Goetz Von Berlichingen  

in Germany and Ambroise Paré in France.  Driven by springs, levers, and gears,  

users can initiate a grasp by the hands via  exaggerated movements of the chest or arm.

Such ways were generally how people  controlled their prosthetic hands back  

then - which was impractical. But  what other methods were available?

In the 1910s Ferdinand Sauerbruch, a famous  German surgeon, created a natural-looking  

prosthetic hand controlled by pins surgically  inserted into the user's forearm. Ouch!

That particular prosthetic hand was ultimately  not used for cost reasons ... plus all the  

infections and inflammation stemming from the  necessary surgical preparations for the stump.

But the hand introduced the idea of "myoelectric  control" - meaning control using electric signals  

generated by the body's muscles. And this  idea remained relevant. Significant work  

was done in the 1940s and 1950s  in Germany and the Soviet Union.

One rather unknown myoelectric Yugoslav  pioneer was Ljudevit Vodovnik of Slovenia,  

who partnered with Case Institute of  Technology in Cleveland to pioneer  

assistive applications. I don't  think he even has a Wikipedia page.

But even these myoelectrically controlled  systems were quite crude. You positioned  

and then commanded them to brainlessly execute a  pre-determined sequence to perform a simple, crude  

action. They were like glorified litter pickers.  Can we imbue them with a little more smarts here?

## From Zero to Max Error

Tomović argued that while the prosthetic hand's  components and circuits had vastly improved  

after World War II, the mathematical  theory to control them lagged behind.

In a 1965 interview with a  US newspaper, Tomović said:

> We are not concerned with  changing the 'hardware';  

rather we are working on introduction of  new mathematics techniques in computer use

He sought to improve these theories by modeling  after how biological systems worked - taking  

in sensory inputs from the outside world  and responding to them in a natural way.

Regarding the artificial hand, he pointed to  how existing systems relied on "zero error".

Such systems worked by monitoring their  own state against an internal reference  

or command signal. The difference  between the two was called "error",  

and the goal was to maintain  the smallest possible error.

So if a robot hand's reference state  says it should be at 30 degrees but  

the actual hand's sensors tell it it is at 28,  

then we seek to reduce that error of  2 degrees in subsequent movements.

Zero error might be helpful for certain  cases like industrial robot arms, but is  

not suitable in life where you encounter  new and different things all the time.

And Tomović also noted that this was  not how biological systems in the world  

functioned. So as an alternative, he proposed  a reversal: A system that maximized "error".

He brought up the scenario of a  hand clasping around something,  

and noticed how hands - even those of  a baby - achieve this by maximizing a  

feedback signal. In other words, the amount  of skin area touching the target object.

He realized that this simple theoretical  model can be leveraged to flexibly handle  

diverse situations. Instead of telling the  hand exactly where to go to clasp something,  

give it a few loose goals and conditions  and let it figure out on its own via a  

"error maximization" feedback loop.

His experiment involved an electrically-powered  skeleton hand - same structure as a  

person's - with a special rubber glove on it. The  glove is conductive. So pressure on the glove can  

produce an electric feedback signal that the  system is incentivized to maximize. Once fully  

clasped, the signal stops rising and the hand  stops applying pressure. Automatic feedback.

Thusly, the artificial gloved hand can  close its fingers around a wide variety  

of differently sized objects without  requiring any special programming for  

each object's individual size and circumstance.

This flow of defining the end goal -  which was to clasp the item - but not  

rigidly defining the transition  states to achieve it might seem  

rather obvious today. But it was a major  leap from prior control theory regimes.

Tomović wrote in the paper’s  conclusion that he and his team  

intended to use these principles to build a  full prosthetic hand and possibly a robot.

## The First Hand

Work on the first Belgrade Hand began  in 1963 and was presented in early 1964.

Powered by an external power source, the  Belgrade Hand was about the same size as a  

normal human hand and mostly looked like  one too. It was attached to the forearm  

and moved into action by a pressure  sensor triggered by the user's biceps.

It had five fingers, the first recorded  robotic hand with that many. The four  

non-thumb fingers' mechanics  operated like the natural finger,  

with three rigid sections  connected with flexible joints.

The thumb however worked differently. It can  only rotate into one of two positions. One,  

outside of the hand and the other inside  the hand just opposite the middle finger.

Building on the biology-inspired  feedback principles we spoke on earlier,  

the hand can grasp things of various  sizes using one of two techniques.

In the first, the thumb rotates to outside of the  hand. When any of the finger pads touch an object,  

the four non-thumb fingers then  curl together around that object  

in a clenching movement. I am somewhat  reminded of those Venus fly traps.

If the object is thick - like  a book or something - then the  

fingers' curling progress is impeded  and the hand ends up in a position  

something akin to a hook. This can  be useful for grabbing door handles.

But if the object is small, then the fingers  hit their limits first. And then the thumb  

rotates inward from its original position to rest  outside the clenched fingers - creating a fist.

The second major grasping movement  that the hand can do was with its  

fingertips. Kind of like an "OK"  symbol with your hands. I imagine  

it might be helpful for picking  up a small thing with precision.

All of these movements were powered by a single  geared motor located partially outside of the  

hand. That motor's force is then distributed using  a complicated set of springs and levers. Pressure  

sensors in the fingers provided the feedback  for the system to determine its movements.

The big deal about the Belgrade Hand was  how it managed to take in and respond to  

feedback from the outside world to do precision  movements. You can make the argument that the  

thing itself had intelligence - taking on  some of the mental burden from the user.

And unlike another pioneering robotic hand -  the MH-1, produced at MIT by Henrich Ernst in  

1962 - the Belgrade Hand's intelligent  capabilities were achieved without the  

use of a large digital computer attached  to it. It seemed to me an analog tool,  

reflecting Tomović's great  experience in the space.

## The Second Hand

A clinical evaluation of this  first hand identified drawbacks.

The first Belgrade Hand's mechanical  complexity made it difficult to manage.  

And its large external power system made  it too heavy and bulky for practical use.

But its capabilities were interesting enough  to receive funding from the US Vocational  

Rehabilitation Administration  in Washington in the mid-1960s.  

The VRA was looking to develop an externally  powered, multi-functional hand-arm system.

So throughout 1966 and 1967, the robotics  team at the Pupin Institute collaborated  

with the VRA and the US National Academy of  Sciences. The partnership allowed Americans  

to travel to Yugoslavia - a rare enough  occurrence to sometimes make the local news.

The second hand operated using the same control  

principles as the first but with tweaks to  the lever system, finger joints and thumb.

The control logic was improved to  offer the user more flexibility and  

predictability regarding how  the hand grasped and opened.

After completion, the hand was  slated for demonstration alongside  

eight other candidates at the Seventh Workshop on  

Upper-Extremity Prosthetics Components  in Santa Monica on July 30-31, 1969.

Unfortunately, this second-generation  prosthetic was also too heavy to be  

fitted onto the tester and  was not selected for use.

A bit later, NASA considered using  the hand as a possible end effector  

for some future space project but  this did not work out as well. But  

the experience evidently brought up  the idea of the hand's use for robots.

## Rise of the Hands

At the start of the 1980s, most robot hands  were just simple two-fingered grippers.

But as the decade progressed, a number of  things happened. Semiconductor microprocessors  

and components shrank in size while also  improving in performance. Combined with  

increased investments from both government  and industry, robot hands rapidly improved.

Two of the most famous robot hands of this  era were Ken Salisbury's three-fingered hand  

and the Utah-MIT Dextrous Hand. The first  became an icon and an actual commercial,  

albeit somewhat expensive product.

The hand's designer, Ken Salisbury,  later went on to join Intuitive Surgical,  

where he contributed to the  famous da Vinci surgical system.

The second hand was a very advanced  five-finger hand that the National  

Science Foundation funded other  universities to purchase and study.

This new environment brought out renewed  interest in the Belgrade Hand and in the  

late 1980s, Tomović teamed up with  long-time friend George Bekey of  

the University of Southern California to  make a new version: The Belgrade/USC Hand.

## The Belgrade/USC Hand

With this robot hand, Tomović and  Bekey wanted to demonstrate the  

validity of a new philosophy of control theory.

The contemporary paradigm of robot locomotion  was a complex, top-down system that receives  

input signals from sensors, processes  those signals, recognizes patterns,  

infers actions in response, and  then plans out motions to execute.

The two argued that this was too complicated  and brittle for practical use - necessitating  

explicit mathematical models of  trajectories and movement and what not.

They instead argued for a philosophy  of control based on stored,  

reflex-like responses triggered by learned  sensory patterns. Such a philosophy is more  

akin to that of living animals and  requires far less complex interplay.

They called it "artificial reflex  control" or ARC. The Belgrade/USC  

Hand was intended to demonstrate that a  robot hand can pick up and grasp items  

of varying size with just ARC principles  rather than complex top-down instructions.

In terms of physical changes, it  uses four separate motors for its  

non-thumb fingers for additional flexibility.

The thumb was made to be fully articulated  with two joints. So it can rotate at an  

axis parallel to the wrist opposite  the index, middle, and ring fingers.

This hand was designed to be mounted  onto a robotic arm called a Puma 560  

at the wrist. An IBM PC/AT computer is  added for higher level tasks: Visually  

identifying the target object,  determining a grasping strategy  

and pre-shaping and positioning the hand.  Its max carrying capacity was five pounds.

Five or six Belgrade/USC Hands were  eventually fabricated and sold to  

university research labs in the US, Germany  and Yugoslavia. But in the commercial space,  

it was too complicated to compete  against simple robot arm grippers.

Bekey tried to get funding for a  third version but couldn't do it.  

He recalls later that a small company  in California tried to commercialize  

the hand but failed to sell more than two or  three, losing a lot of money in the process.

Unfortunately, Yugoslavia entered turbulent times  soon after the Belgrade/USC Hand's presentation  

in 1990. Tomović retreated from the public -  passing away later in 2001 at the age of 81.

Bekey continued studying biologically  inspired robot control throughout the 1990s.  

He passed last year at the age of 96.

## Conclusion In a 2008 interview with Wired Magazine,

George Bekey says that a robot hand  today would not need five fingers.

Hands like the Salisbury have largely  proven that three fingers are good enough.

But the Belgrade Hand nevertheless stands as  a pioneering achievement in robotics history,  

frequently mentioned in papers  for its groundbreaking nature.

And it and its creators are celebrated today as a  sign of technological progress in old Yugoslavia.

Even these remembrances largely skip over  what I think was Tomović's north star:  

His biologically-inspired theory  of controlling robots via local  

reflex actions rather than complex  top-down systems. A radical idea.  

One that continues to fascinate people  and spur on future research to this day.