How Gentle Is Gentle Enough? Building Robots That Know Where to Stop
Imagine handing a surgeon a scalpel. You don’t need to place it in an exact spot down to the centimeter—you would just need to get it close and, above all, not stab anyone by accident. That distinction between precision and safety sits at the center of CISE Faculty Affiliate and Assistant Professor (ME, SE) Andrew Sabelhaus’ research on soft robots.
Sabelhaus directs BU’s Soft Robotics Control Lab, where his team builds robots from rubber and other pliable materials, then develops the mathematics that lets them move on their own. Unlike the rigid metal robots that dominate factory floors, soft robots are meant to touch things, bump into people, and operate in environments that can’t be fully planned out in advance—which raises questions that don’t come up on an assembly line.
“There are really deep questions about what it means for that robot to actually do its task right,” Sabelhaus said, “or what it means to even do a task in the first place.” What does “safe” mean for a robot helping a doctor or lifting a patient out of bed? How much contact is too much? Sabelhaus’ lab is trying to turn these doubtful, human questions into precise engineering answers.
The lab’s working hypothesis is that there is no single, universally accepted definition of “safety” or “success.” However, there is a language for specifying an explanation, case by case. Using a mathematical framework called temporal logic, engineers can encode rules like “force must stay under this threshold” or “this action must happen within this time limit.” From there, Sabelhaus’ team works out how to translate that logic into an actual robot’s motion—essentially building an autonomous system that inflates or moves in exactly the way needed to satisfy the specification.
It is a distinctly control-oriented approach that gives the lab its name. Other soft robotics labs, including several at BU, build robots that are far more mechanically sophisticated. However, Sabelhaus is more interested in how well the robot can make decisions. A complex soft robot with limited sensing and no real feedback logic, he explained, isn’t much different from a factory arm: great at repeating the same motion, but unable to adapt when something changes. His lab instead tried to build simple robots, but well-instrumented enough to make nuanced decisions grounded in solid math.
The project Sabelhaus is most excited about right now is safety-verified control—engineering a way to mathematically guarantee that a soft robot won’t apply too much force, even while it’s actively touching something. Last year, his lab demonstrated the first hardware example of a verifiably safe soft robot making contact with its environment, with provable bounds on how hard it pushes.
The next step is applying that idea to real tasks, and one of Sabelhaus’ favorite examples is agricultural: teaching a robot to pick a berry. Success means the berry actually detaches from the bush—but squeeze too hard and the berry is crushed. It’s a small, well-defined problem, which is exactly the point. Sabelhaus is cautious about the idea of robots making sweeping autonomous decisions in high-stakes settings like surgery. “Everything that my group is working on is really taking really small pieces of the problem,” he said—mapping a narrow slice of human expertise into something precise enough for a robot to execute, before ever thinking about bigger decisions.
When asked whether large language models and generative AI might eventually reshape this work, Sabelhaus was candid about the uncertainty. Safety-based robot control has relied on similar mathematical foundations for decades, and the rise of generative AI has “turned all that on its head.” Though exactly how the two will merge, if they do, remains an open question. In Sabelhaus’s view, the largest defining challenge in soft robot control isn’t hardware or algorithms; it is translation. Humans can describe what should or shouldn’t happen using ordinary language, but converting that intuition into a precise engineering specification—how hard is “too hard,” which finger should make contact, what the time limit should be—remains largely unsolved. Bridging that gap between human intuition and formal specification is a problem his field will eventually have to confront directly.
This semester, Sabelhaus is focused on teaching two courses that might look different on paper but share the same mathematical backbone: EK 103 (Computational Linear Algebra), for first-year undergraduates, and ME 501 (Dynamical Systems Theory), for first-year graduate students. In both, he enjoys showing students the classical version of a problem—like positioning a rigid robot arm in space, a technique that’s been well understood for decades—before posing the harder question: what happens when the robot is soft and touches a person unexpectedly during a medical procedure? For Sabelhaus, that’s where the real research begins, using tools that started in an introductory classroom and ended up back in the lab.
Sabelhaus held a NASA Space Technology Research Fellowship with NASA Ames Research Center, then went on to an Intelligence Community Postdoctoral Research Fellowship at Carnegie Mellon University before arriving at BU. Both experiences shaped the kinds of questions he considers worth pursuing, pushing him to think about broader societal priorities rather than just what he personally finds interesting. That instinct to question his own assumptions feeds directly into his research on defining safety and success—an exercise in staying open to the possibility that there isn’t one obvious right answer. Sabelhaus received the NSF CAREER Award in 2024, and in 2026 he was named Professor of the Year in Mechanical Engineering at BU.