A quadruped robot built to mirror the natural motion and stability of a running dog has achieved something unusual in robotics: diverse, animal-like movement using only four actuators to control twelve joints. The machine, called PAWS—short for Passive Automata With Synergies—uses mechanical intelligence borrowed from biology to handle disturbances, vary its gait, and even overcome obstacles without needing constant active control.

Most legged robots today rely on heavy computation and numerous motors. Each joint typically gets its own actuator, and complex algorithms coordinate them all in real time. The approach works, but it demands dense sensor networks, fast feedback loops, and significant processing power. Animals do things differently. Their nervous systems reduce complexity by grouping muscles into coordinated patterns called motor synergies, while passive properties like springy tendons and compliant joints handle much of the interaction with terrain.

PAWS brings both ideas into hardware. The robot extracts coordination patterns from motion capture data of dogs, then translates those patterns into a mechanical design where tendons route across multiple joints and springs provide tunable stiffness. The result is a machine that exhibits emergent behaviors, moving in ways that look strikingly biological even when no motors are running.

Extracting Movement Patterns from Animal Data

The design begins with over 147,000 poses captured from a dog performing a range of behaviors: walking, trotting, galloping, sitting, jumping, lying down. This dataset reflects the full spectrum of natural motion. The research team applied principal component analysis to identify which joint movements tend to occur together. The technique revealed that just four coordinated patterns, or synergies, could explain roughly 88 percent of the variance across all twelve leg joints when all four legs were considered together.

Breaking the body into left and right halves, each with a fore and hindlimb, the team found that two synergies per side could still capture around 79 percent of motion variance. That level of coverage is high enough to replicate most natural behaviors while dramatically reducing the number of independent controls needed. Instead of twelve motors, the design requires only four: two per side, each pulling a tendon that implements one synergy.

Each synergy represents a specific ratio of joint activations. When a tendon is pulled, it simultaneously moves multiple joints in a coordinated way that mirrors how the dog's muscles fire during real movement. This approach does more than save on actuators. It embeds biological coordination directly into the mechanical structure, allowing the robot to move through postures that resemble those seen in the animal data even without explicit joint-level control.

Building Compliance Into the Body

The robot consists of four legs, each with three joints, arranged symmetrically. Links are machined from aluminum, and each joint houses a torsional spring made from two linear springs acting antagonistically on a pulley. The stiffness at each joint can be tuned by changing the pulley radius, which scales quadratically. This modular design allows rapid customization to match the stiffness values identified during optimization.

An automated pipeline takes the extracted synergies and determines the best tendon routing, pulley sizes, and joint stiffness to reproduce those synergies mechanically. The optimization minimizes the difference between desired synergy activation and the motion the physical system actually produces. The resulting design shows high stiffness in the front foot and rear hip, with lower stiffness in the rear foot and front hip, creating an inverse stiffness distribution between fore and hindlimbs.

This distribution is not arbitrary. It reflects the interaction forces animals experience during locomotion. During the stance phase, when a paw is on the ground, the system is stiffer perpendicular to the trajectory to support body weight. During the swing phase, when the leg is airborne, compliance becomes parallel to the motion, allowing the limb to react gently to unexpected contact.

The tendons themselves are thin metal cables that wrap around pulleys positioned at each joint. Each motor pulls a single tendon that spans multiple joints, coupling them together. This mechanical coupling means that when one leg encounters an obstacle or perturbation, the other leg responds passively through the shared tendon, without needing sensor feedback or active correction.

Passive Gaits Emerge From Treadmill Interaction

One of the most striking demonstrations involves placing the robot on a moving treadmill without connecting the motors to the tendons. The four-bar linkage stabilizes the body in the sagittal plane, and a counterweight partially offsets gravity. As the treadmill belt moves, the robot's feet interact with the surface, storing and releasing energy in the springs and tendons. The result is a periodic, galloping-like gait that emerges purely from the passive dynamics.

Video analysis shows the robot rising and falling rhythmically, with coordinated movement between fore and hindlimbs. The pattern resembles a typical running gait in dogs, complete with alternating stance and flight phases. As treadmill speed increases from one to six kilometers per hour, the gait adapts. The forelimb extends further forward at higher speeds, increasing stride length. The hindlimb amplitude grows until around four kilometers per hour, then plateaus. Meanwhile, the proportion of time spent in flight increases, and the stance phase shortens—trends that match biological locomotion at higher speeds.

This variation in response to external energy input shows that the passive system is not locked into a single pattern. The directional stiffness and leg coupling allow the robot to shift between distinct stable behaviors depending on environmental conditions, all without active control.

Robustness to Disturbances

Robustness is a hallmark of animal locomotion. When a dog steps on uneven ground or is nudged mid-stride, it recovers quickly without needing to plan a corrective response. The research team tested whether PAWS exhibits similar resilience.

In one experiment, an external stick was used to perturb the forelimb, body, or hindlimb while the robot galloped passively on the treadmill at four kilometers per hour. Cameras tracked the paw trajectories before, during, and after the disturbance. In all cases, the gait rapidly returned to its previous limit cycle. When the forelimb was displaced, a small coupled response appeared in the hindlimb, but both legs resumed their steady trajectory within one or two cycles. Disturbances to the body affected both legs but were similarly absorbed. The hindlimb, which is stiffer at the hip, took about four cycles to return to equilibrium after being perturbed, reflecting the greater energy stored in that stiffer joint.

The team also placed obstacles of varying heights on the treadmill belt. At twenty millimeters—roughly one fifth the robot's leg height—the passive system cleared obstacles with complete success. At forty millimeters, success dropped to around seventy percent, and at sixty millimeters, to sixty-five percent. The fact that any obstacles are cleared without active sensing or control suggests that the mechanical design alone provides substantial adaptive capacity. Beyond a certain height, active control becomes necessary, mirroring the way animals combine passive mechanics with neural intervention.

Active Control Builds on Passive Foundation

While passive gaits demonstrate the power of embodied intelligence, the robot becomes fully capable when the four motors are engaged. Using inverse kinematics in synergy space, the team can generate motor commands that drive the robot through a variety of biologically inspired behaviors: walking, sitting, crouching, standing, leaning forward, and jumping.

For walking, the desired paw trajectories are taken from the dog motion capture data and scaled to the robot's size. The inverse kinematics solver finds the synergy activations that best match the Cartesian foot positions. Despite optimizing only for foot placement, the resulting joint angle evolution closely resembles the animal data. This correspondence occurs because the synergy basis implicitly guides the redundant leg system toward biologically relevant postures. In other words, the mechanical structure helps the robot move like a dog without explicit joint-level programming.

When the actuated robot walks on the treadmill at the same speed as the passive version, the gait amplitude increases both horizontally and vertically. The robot can also clear higher obstacles. Yet the passive robustness is not lost. When a three kilogram obstacle is placed on the belt, the actuated gait is perturbed but recovers within a few cycles, showing that underactuation and distributed compliance preserve disturbance rejection even during active control.

The robot can also transition between behaviors. In one demonstration, it gallops passively on the treadmill, then motors fire briefly during hindlimb contact to trigger a jump. After landing, the passive gallop resumes. This blending of passive and active modes illustrates how a well-designed body can scaffold complex behaviors with minimal control input.

Mechanical Couplings Create Coordinated Responses

The synergistic tendon routing does more than reduce the number of motors. It creates mechanical couplings between legs that produce coordinated responses to external forces. When the forelimb is pushed forward, the hindlimb moves backward and downward. When the hindlimb is pushed forward, the forelimb shifts horizontally. These coupled motions depend on both the robot's configuration and the direction of the applied force.

Such coupling can assist with obstacle navigation. If the forelimb encounters an obstacle, the coupled hindlimb response helps maintain stability. If the hindlimb is disturbed, the forelimb adjusts accordingly. These are embodied compensatory strategies—mechanical reflexes built into the structure rather than computed by a controller.

Experiments using a robotic arm to displace one paw while the other was tracked showed that coupling strength varies across the workspace. This variation arises from the configuration-dependent stiffness of the tendon and spring system. The stiffness ellipsoids measured at different points in the leg's reach show that impedance is higher and perpendicular to the trajectory at the extremes of motion, where the leg supports weight, and lower in the center, where compliance aids swing.

Implications for Robot Design and Control

The results challenge the prevailing approach to legged robotics, which often treats the body as a platform for computation rather than an active participant in control. PAWS shows that by carefully designing passive mechanical properties and actuation synergies, a robot can achieve diverse, robust, and natural-looking locomotion with far fewer actuators and sensors than conventional designs require.

This does not mean computation is unnecessary. High-level planning, terrain adaptation, and complex tasks still benefit from active control. But offloading low-level stability and disturbance rejection to the body reduces the burden on the controller and may improve energy efficiency and response speed. For tasks like navigating rough terrain or recovering from unexpected contact, passive mechanics can provide an immediate, zero-latency response that active control cannot match.

The synergy-based design also opens a path toward more versatile underactuated systems. Hands have successfully used synergies to achieve complex grasping with minimal actuation. PAWS extends this principle to full-body locomotion, where coordination must span multiple limbs and respond to dynamic environmental interactions.

Future work could explore how variable stiffness might be incorporated, allowing the robot to tune its compliance in real time to suit different terrains or tasks. Another direction involves scaling the approach to more complex environments or integrating it with learning-based controllers that adapt synergy activations based on experience.

A New Paradigm for Machine Intelligence

PAWS represents a shift toward machine physical intelligence, where the structure of the robot itself embodies solutions to control problems. Rather than solving for stability and coordination solely through software, the design encodes these properties in springs, tendons, and mechanical couplings. The result is a machine that behaves less like a programmed automaton and more like a biological organism responding to its environment.

The work draws on decades of neuroscience research showing that animals use hierarchical motor control, with the nervous system activating groups of muscles rather than individual fibers. By translating this principle into hardware, the researchers have created a robot that mirrors not just the appearance of animal motion but the underlying mechanical and neural organization that makes that motion possible.

For robotics, the implications are practical. Reducing the number of actuators lowers weight, cost, and power consumption. Simplifying control reduces computational overhead. Embedding robustness in the body makes the system more reliable in unpredictable settings. For biology, the work provides a physical model that can test hypotheses about how synergies and passive properties contribute to real animal locomotion.

In the end, PAWS is more than a clever engineering solution. It is a demonstration that rethinking the boundary between body and brain—between passive structure and active control—can unlock new forms of robotic capability. The dog-like quadruped that gallops on a treadmill without motors or walks, jumps, and crouches with just four actuators is not mimicking intelligence. It is embodying it.

Credit & Disclaimer: This article is a popular science summary written to make peer-reviewed research accessible to a broad audience. All scientific facts, findings, and conclusions presented here are drawn directly and accurately from the original research paper. Readers are strongly encouraged to consult the full research article for complete data, methodologies, and scientific detail. The article can be accessed through https://doi.org/10.1038/s42256-025-00988-x