Muscles were sensing external forces, not just executing commands
At Peking University, scientists have discovered that fish muscles do not merely propel — they perceive, reading the surrounding water as much as they drive the body through it. By decoding the electrical signals of swimming koi and carp, researchers uncovered a sensorimotor logic so transferable that a model trained on living fish could guide a robot it had never encountered. This finding invites us to reconsider where intelligence resides in a body, and what machines might become when they are taught not just to imitate life's forms, but to inherit its deeper principles.
- A foundational assumption in biology and robotics has quietly collapsed: fish muscles are not just engines, they are sensors that read the water around them.
- The discovery emerged from a striking reversal — in turbulent conditions, muscles fired after the body moved, revealing that fish respond to deformations imposed by the environment, not just commands from within.
- Researchers built a mathematical model of fish sensorimotor control so precise it captured timing delays, response strength, and tail oscillation frequency from real electromyography data.
- When applied to a robotic fish without any retraining, the fish-derived model outperformed systems trained directly on robot data — a result that unsettles the boundary between biological and mechanical intelligence.
- The field of underwater robotics now faces a new design horizon: machines that don't mimic the shape of fish, but inherit the sensory logic embedded in their muscles.
Scientists at Peking University spent months recording the electrical signals firing through fish muscles as koi and carp swam through both smooth and turbulent water. Using electromyography alongside high-resolution video, they trained a deep neural network to translate muscle signals into body position — and what the data revealed was unexpected.
In calm water, muscle activation preceded movement, the familiar pattern of an animal driving itself forward. But in turbulent conditions, the sequence reversed: muscles fired after the body had already shifted, meaning the fish were sensing deformations imposed by the water and responding to them. The muscle, it turned out, was not just an engine — it was a sensor reading the surrounding environment in real time.
From this insight, the team built a mathematical model capturing the full relationship between electrical signals and tail motion, including timing, response strength, and oscillation frequency. The true test came when they applied this model — trained entirely on living fish — to a robotic fish, with no retraining for the mechanical system. It predicted the robot's tail movement more accurately than a model trained on robot data alone.
The implications extend well beyond any single laboratory. Underwater robots built on these principles could navigate currents and disturbances the way living fish do, responding to the water rather than merely pushing through it. The research also opens new possibilities for biological telemetry and for understanding how animals process fluid environments. What began as a study of muscle signals has become a blueprint for a new generation of machines — ones that don't just wear the shape of life, but carry something of its intelligence.
Scientists at Peking University have spent months watching fish move through water, recording the electrical signals that fire through their muscles as they swim. What they discovered challenges a basic assumption: fish muscles do far more than propel bodies forward. They sense. They read the water. And now, researchers believe those capabilities could transform how we build robots meant to navigate underwater environments.
The work began with electromyography—a technique for measuring the electrical activity inside muscle tissue. The team recorded these signals from koi and carp as the fish swam through different conditions: smooth, laminar flows and chaotic vortex sheets. Simultaneously, they captured video of each fish's body position and movement. A deep neural network learned to map the muscle signals directly to joint angles, essentially teaching itself to read what the fish's muscles were saying about how its body was positioned.
What emerged from this data was unexpected. The muscle signals contained information about the water itself, not just the fish's own motion. When researchers looked at the sequence of events in smooth water, muscle activation came first, followed by body movement—the expected pattern for an animal pushing itself forward. But in turbulent conditions, the order flipped. Muscles activated after the body had already moved, suggesting something else was happening. The fish were sensing deformations imposed on their body by the surrounding water and responding to them. Their muscles weren't just engines; they were sensors.
In a second phase, the researchers built a mathematical model linking muscle activity directly to tail movement. This model captured the relationship between electrical signals and physical motion, including the timing delays, the strength of response, and the natural frequency at which the tail oscillated. It was a distillation of how a real fish's sensorimotor system actually worked.
Then came the test that mattered most. The researchers took this model—trained entirely on fish data—and applied it to a robotic fish. Without any retraining specific to the robot, the model could predict how the robot's tail would move. It worked better than a deep learning system trained directly on robot data, suggesting that the underlying principles of fish muscle control were genuinely transferable to mechanical systems.
The implications ripple outward. Underwater robots designed with these principles could navigate complex environments more effectively, responding to currents and disturbances the way living fish do. The work opens doors to better biological telemetry, to deeper understanding of how animals move through fluid environments, and to a new generation of machines that don't just mimic the shape of fish but capture something of their actual sensorimotor logic. The muscle, it turns out, is smarter than we gave it credit for.
Notable Quotes
Muscle signals can decode motion and flow, revealing the sensing built into fish and guiding better underwater systems based on sensorimotor dynamics— Peking University research findings
The Hearth Conversation Another angle on the story
So the researchers recorded electrical signals from fish muscles. What made them look there in the first place?
Fish have been solving the problem of underwater movement for millions of years. If you want to build a robot that moves like a fish, you need to understand not just the skeleton and fins, but the actual control signals—the electrical commands that coordinate everything.
And they found the muscles were doing something unexpected?
Yes. In turbulent water, the muscle activation came after the body had already moved. That's not a propulsion signal—that's a response to something the water did to the fish. The muscles were sensing external forces.
How does that change robot design?
Instead of a robot that just executes a pre-programmed swimming pattern, you could build one that reads its environment through its own actuators and adjusts in real time. It becomes reactive, adaptive.
Did the model they built actually work on the robot?
It worked better than expected. A model trained only on fish data predicted robot tail motion more accurately than a system trained directly on the robot. That's the real proof—the principles transfer.
What comes next?
Underwater exploration, search and rescue, environmental monitoring. Any task where you need a machine that can handle unpredictable currents and complex flows. The fish have been teaching us, and now we're finally listening.