The robot doesn't have to ask how much force it's applying. It already knows.
At Seoul National University, engineers have woven sensing and movement into a single artificial muscle — a quiet but significant step toward machines that don't merely act upon the world, but feel it as they do. For decades, robotics has treated motion and perception as separate concerns, a division that biological life never accepted. This new structure, threaded with liquid-metal channels inside a shape-shifting elastomer, suggests that the boundary between tool and organism may be less fixed than we assumed.
- The core tension is old and stubborn: robots that handle delicate tasks have always needed two separate systems — one to move, one to sense — creating wiring complexity, processing lag, and more points of failure.
- SNU's artificial muscle collapses that division into a single component, with one liquid-metal channel driving contraction through heat and another reading force and deformation in real time.
- Testing on robotic fingers showed the integrated muscle could identify object stiffness and size while gripping gently, and paired opposing muscles achieved faster, self-correcting control without waiting for external feedback.
- Heat buildup during repeated movement causes force drift and tracking errors — the muscle loses accuracy as it tires, and the stretch estimation model still needs real-world refinement.
- The team is targeting thinner materials, built-in cooling channels, and solid-state Peltier modules to dissipate heat faster — treating what remains as an engineering problem, not a fundamental barrier.
Engineers at Seoul National University have built an artificial muscle that does what biological tissue does naturally: it moves and senses at the same time. Threaded with liquid-metal channels inside a liquid-crystal elastomer, the device uses one channel to generate contraction through electrical heating and another to measure force and deformation in real time. The researchers call this combination physical intelligence — a single structure that both acts and knows.
The problem driving the work is familiar to anyone who has watched a robot fumble with a fragile object. Delicate manipulation requires motion and sensing to work in concert, but conventional robots keep them separate — actuators on one side, external sensors bolted on the other. That separation means more wiring, more processing, slower reactions, and more places for things to go wrong. Biological muscle never accepted this division: a human muscle fiber contracts while simultaneously reporting tension and load back to the nervous system, and the brain adjusts instantly.
The SNU team designed their structure to mirror that integration, connecting two types of liquid-crystal elastomer in series to mimic the relationship between tendons and muscle fibers. When tested on robotic fingers and grippers, the results were telling. The artificial muscles could pick up objects while simultaneously gauging their stiffness and size. Pairing two muscles in opposition — as human limbs do — produced faster, more precise control, with the system correcting its own errors without waiting for outside feedback.
The technology is not yet ready for deployment. Heat accumulating during repeated movement causes force drift, and sudden changes in target position create tracking errors. The stretch estimation model needs further refinement before it can be trusted in real applications. The team's proposed fixes — thinner materials, cooling channels, integrated Peltier modules — are engineering challenges rather than conceptual dead ends. The direction is what matters: toward robots that don't carry sensors as accessories, but carry sensing as a property of their own bodies.
At Seoul National University, engineers have built something that works like muscle tissue—contracting on command while simultaneously sensing what it's doing. The device uses liquid-metal channels threaded through a liquid-crystal elastomer, a material that responds to electrical current by changing shape. One channel heats up and pulls the structure tight. The other measures the force and stretch in real time. Together, they create what the researchers call physical intelligence: a single component that both moves and knows.
The problem they were solving is straightforward but stubborn. Robots that need to handle delicate tasks—picking up an egg, grasping a hand, assembling something fragile—require two separate systems working in tandem. One system makes the robot move. Another system, usually external sensors bolted onto the structure, tells the robot what it's touching and how much force it's applying. This separation creates complexity. More sensors mean more wiring, more processing power, more places for things to fail. It also means the robot can't react as quickly or as naturally as a living thing would.
Biological muscle doesn't work that way. A human muscle fiber contracts and simultaneously feeds information back to the nervous system about tension, length, and load. The brain receives this data instantly and adjusts. The system is integrated. SNU's team designed their artificial muscle to mimic this integration. They connected two types of liquid-crystal elastomer—one isotropic, one nematic—in series, creating a structure that mimics how tendons and muscle fibers work together in the body. The liquid-metal channels embedded inside handle both the actuation and the sensing.
When the researchers tested the system on robotic fingers and grippers, the results showed what this integration could do. The artificial muscles could pick up objects gently while simultaneously identifying how stiff those objects were and how large. By pairing two muscles to work against each other—mimicking the opposing muscle pairs in human limbs—the team achieved faster, more precise control. The finger could contract and relax with better accuracy, and the system could correct its own errors in real time without waiting for external feedback.
But the technology isn't yet ready for deployment in the next generation of humanoid robots. Heat buildup inside the artificial muscle during repeated movements causes what researchers call force drift—the muscle gradually loses accuracy as it tires. Sudden changes in movement targets create tracking errors. The stretch estimation model, built from experimental data, needs refinement before it can be trusted in real-world applications. The team has identified solutions: thinner materials to dissipate heat faster, built-in cooling channels, or Peltier modules—solid-state cooling devices—integrated into the muscle itself. Faster cooling would improve both response speed and sensing reliability.
What matters here is the direction. Humanoid robots and assistive devices are coming. They'll need to move like humans move and sense like humans sense. The conventional approach—bolting sensors onto actuators—won't scale. It's too slow, too power-hungry, too prone to failure. The SNU team has shown that embedding sensing directly into the actuator itself is possible. The challenges that remain are engineering problems, not fundamental barriers. As the technology matures, robots will become less like machines with sensors attached and more like organisms with integrated nervous systems.
Citas Notables
The system combines sensing and movement in a single structure, allowing the artificial muscle to monitor its own condition in real time without relying on external sensors.— Seoul National University researchers
La Conversación del Hearth Otra perspectiva de la historia
Why does it matter that the sensing and movement are in the same structure? Couldn't you just use better external sensors?
You could, but you'd be fighting physics. External sensors add lag—the robot moves, then waits for the sensor to report back, then adjusts. A living thing doesn't work that way. The feedback is instantaneous. Also, every sensor is another component to power, another point of failure, another layer of complexity in the control system.
So this artificial muscle is faster because it knows what it's doing as it does it?
Exactly. It's contracting and measuring simultaneously. The robot doesn't have to ask "how much force am I applying?" It already knows. That changes everything about how naturally it can interact with the world.
The article mentions heat buildup as a problem. Why does heat matter so much?
Heat causes the material to drift. The muscle loses calibration. After repeated movements, it's no longer as accurate. It's like a musician's instrument going out of tune. You can still play, but the precision is gone.
Can they just make it cooler?
That's the plan. Thinner materials, cooling channels built in, maybe active cooling. But it's a trade-off. You want the muscle to be strong and responsive, but you also need it to stay cool. Finding that balance is the engineering work ahead.
When do you think we'll see this in actual robots?
The core idea is proven. The remaining problems are solvable. I'd guess a few years before you see this in specialized applications—medical robots, delicate assembly work. Humanoid robots that move like humans? That's further out, but this is a necessary step.