AINEWS
A soft silicone hand driven by air pressure, with fewer motors than joints and the same control command for everything it picks up, just won a once a decade lifetime achievement award from the field's top academic venue.
A soft silicone hand curls around a raw egg and lifts it without breaking the shell. The same hand, running the same single control command, wraps around a block of tofu and a wobbly toy without crushing either. There is no camera-based reconstruction of the object, no model trained on millions of hours of robot video, and no force-torque math locking the fingers in place. The hand is air-driven, has fewer motors than joints, and decides how to grasp each object by what it is made of.
That hand is the RBO Hand 3, and the 2014 paper that introduced its predecessor just won the Test of Time Award at the 2026 Robotics: Science and Systems conference (RSS 2026), held July 15 in Sydney according to on-site coverage. The award goes once a decade to a paper that has aged into a classic. The reason this one won is the part the trillion-dollar embodied-AI push is now trying to relearn from scratch: a body can do the computation a model has to be trained to perform.
The paper is "A Novel Type of Compliant Underactuated Robotic Hand for Dexterous Grasping," presented at RSS 2014 in Berkeley by Raphael Deimel and Oliver Brock, then at the Technische Universität Berlin. Its journal extension appeared in IJRR in 2016. Its thesis, in three words, was "morphology is computation." Translation: a hand's physical shape, the way it gives way on contact (compliance), and the way it interacts with the object and the world (environmental constraints) can substitute for the control math and the learned perception pipeline that a more rigid, sensor-heavy hand would need.
The 2014 design made the substitution literal. Each finger has two embedded helical threads that stop it from ballooning outward, and a fabric liner at the base that bends but cannot stretch. Inflate the air bladder inside the finger and it curls; deflate it and it relaxes. The whole hand is underactuated, which in plain English means it has fewer motors than joints, so the missing motors are replaced by the physics of the rubber. Press against a hard edge and the finger stiffens; press against a soft one and it conforms. The same control command produces a different grasp on every object, and the control system does not have to know which object it is holding.
That last point is the part the current roadmaps are having trouble with. The dominant 2025-2026 bet in embodied AI is to scale: bigger foundation models, more training data, more hours of robot video, more compute. The premise is that dexterity will fall out of scale the way language modeling did. The RSS 2026 award is the field's most prestigious backdated endorsement of a different premise: if you build the hand right, the model becomes simpler, the sensor suite gets smaller, and the compute bill collapses.
The dollar contrast is jarring, even when it is a media framing rather than a paper result. According to on-site coverage by Lei Feng Wang's AI Tech Review, when the RBO Hand breaks, a few yuan of glue is enough to put it back together, a sum that lands in cents rather than dollars. The hand sits against a generation of dexterous manipulators whose sensor stacks, GPUs, and training pipelines run into the tens of thousands of US dollars. The contrast is not the point of the paper; the paper's contribution is the mechanism. But the award citation is the field admitting, twelve years later, that the mechanism aged well.
The historical arc helps explain why. Soft and underactuated grasping did not start in 2014. Researchers were building envelope-style grippers in the late 1970s, the University of Pisa and the Italian Institute of Technology iterated compliant hands through the 2000s, and a Harvard group published a widely cited soft silicone gripper in 2011. The 2014 RBO paper is the one that pulled those threads into a single design with a clear thesis about why the body, not the controller, is doing the work. The author's bibliography on DBLP traces the follow-on work.
Oliver Brock, his co-author, accepted the award in person. Raphael Deimel, joining by video from Austria, said: "We were not chasing wealth, we were chasing something that would leave a mark in the world." It was a sentence spoken minutes after a 2014 soft hand received a lifetime-achievement award from the same field that is currently building the most expensive embodied-AI systems in history.
Two honest-size caveats. The morphology-as-computation claim is strongest on constrained grasps where the environment cooperates; general-purpose manipulation in clutter is still an open problem that no compliant hand has solved alone. And the dollar contrast is reporter framing, not paper data, so it should be read as a useful frame rather than a benchmark. Neither caveat weakens the award. The paper opened a design direction the field has spent twelve years extending, and the 2026 Test of Time Award is the field saying it noticed.
The next watch item is whether the embodied-AI roadmaps absorb the lesson or build past it. If the trillion-dollar bet is right, scale alone will produce dexterity and the soft hand will be remembered as a curiosity. If the 2014 paper is right, the next generation of dexterous hands will look more like the RBO Hand 3 and less like a server rack wearing a glove, and the cost curve will follow the glue.