What a robotic hand can teach us about intelligence – SCIoI spokesperson Oliver Brock receives Test of Time Award (RSS)

The Test of Time Award is an unusual kind of scientific recognition, because it doesn’t look at the novelty of a scientific contribution, but asks what remained important afterwards. Which ideas changed the way researchers thought about a problem? Which ones opened up new directions? Which ones still matter when the excitement of novelty has long passed?

For us at Science of Intelligence, this particular award touches on a question that has accompanied our work for years: How much of intelligent behavior really has to be controlled?

When less control creates more possibility

Traditional robotic hands often approach dexterity through precision. More joints, more actuators, more sensing, more control. If you want a machine to produce a particular movement, the intuitive solution is to give it the ability to specify that movement as exactly as possible. Soft and compliant robotic systems were not new at the time, but the RBO Hand 2 pushed this approach much further, making compliance central to how the hand manipulated objects.

It was soft, pneumatically actuated, highly compliant, and underactuated. Instead of controlling every possible movement of every part of the hand, its design allowed the physical structure itself to respond to the object it encountered.

What was much less clear was what this would mean for dexterity. Wouldn’t all of this compliance make precise, dexterous grasping harder?

Raphael and Oliver put the problem plainly in the original paper:

“Underactuation and passive compliance seem to render dexterous grasping difficult or even impossible. The experiments performed with our novel hand indicate otherwise.”

And they did. The RBO Hand 2 could perform 31 of the 33 grasp types described in the Feix taxonomy of human grasping. Its opposable thumb also achieved seven out of eight positions in the Kapandji test, a test originally developed to assess human thumb dexterity. What made this particularly interesting was not the fact that the hand could grasp many different objects but how it did so.

Let the body do some of the thinking

The hand had seven actuators, grouped into only four actuation channels for the grasping experiments. And yet the final postures it could produce were considerably richer than those four dimensions of control would suggest. The “missing” complexity came from the interaction between hand and object. When the soft fingers make contact with an object, they deform and wrap around its surface as they adapt to differences in shape and position. A relatively coarse control command can therefore produce a much more specific final grasp because the mechanics of the hand help complete the movement.

This was one of the central arguments of the paper: Raphael and Oliver found that the dimensionality of the resulting grasp postures was greater than the dimensionality of the hand’s control. The idea that grasping behavior can have a lower-dimensional control structure was not new and had already been explored in human hands, notably in work by Marco Santello and others. What Raphael and Oliver showed was how this could play out in a highly compliant robotic hand. Their interpretation was that compliant interaction with the object generated part of this additional variability. In other words, while the body was executing a command, it was also participating in the solution.

This seems like a simple shift in perspective, but it resonates strongly with how we think about intelligence at SCIoI. Intelligent behavior does not necessarily emerge from a controller specifying every detail in advance. Instead, complexity can sometimes be reduced by exploiting the structure of the body, the environment, and especially the interaction between them. Simply put: A hand does not need to calculate every millimeter of an object’s geometry if its fingers can adapt to it.

A different idea of dexterity

The RBO Hand 2 was an exercise in making robots softer, but its design, more importantly, challenged the idea that robustness and dexterity necessarily pull in opposite directions. Compliance had already made robotic grasping more forgiving. A soft hand can tolerate uncertainty in where an object is positioned and can also absorb impacts. It can adjust when reality does not quite match the model.

But the RBO Hand 2 showed that these qualities did not have to come at the expense of versatility. On the contrary, the authors argued that passive compliance might actually help produce dexterous behavior. They were careful about how far they took that conclusion. Some of the paper’s analysis relied on an assumption that the dimensionality of human and robotic grasp postures was comparable, and the authors explicitly noted that stronger evidence would be needed before drawing definitive conclusions.

That caution was important, but so was the direction the work pointed toward: Perhaps sophisticated behavior does not always require sophisticated control. Perhaps part of what appears to be intelligence at the level of behavior can emerge from the right relationship between morphology, control, and environment.

From RBO Hand 2 to the questions we ask today

This idea did not end with this particular hand but became part of a longer trajectory in Oliver’s research on soft and embodied robotic manipulation, from the first RBO Hand through the RBO Hand 2 and RBO Hand 3, which was mainly worked on by Steffen Puhlmann from Oliver’s Robotics and Biology Laboratory (RBO), to current prototypes. And the questions behind it became part of something larger too.

Its fingers adapt instead of requiring the controller to predict every contact. Its material properties simplify a problem that could otherwise demand much more computation. The physical interaction between hand and object becomes part of the mechanism that produces the behavior.

This is exactly the kind of relationship that interests us at SCIoI: not intelligence as something contained in an algorithm, a brain, or a machine, but intelligence as something that emerges through interaction.

What the Test of Time tells us

There is something fitting about receiving a Test of Time Award for this work. Science often rewards novelty, and understandably so. But whether an idea is truly useful sometimes only becomes visible much later: when other researchers have had time to build on it, challenge it, rethink it, or discover that the question it raised has become even more relevant.

The RSS Test of Time Award was created precisely for that kind of reflection. It recognizes papers published at least ten years earlier that changed how researchers think about problems, introduced new research directions, or pioneered new approaches to robotic design.

For the RBO Hand 2, the enduring idea is that intelligence can emerge when we stop trying to control everything. And for SCIoI, that idea reaches far beyond robotic hands. It speaks directly to how we think about embodied intelligence: how biological and artificial systems make use of their bodies, their environments, and the physics of the world to turn limited control into rich and adaptable behavior. And perhaps this is what makes a scientific idea pass the test of time: Not that the original question has been settled, but that we are still asking it.