In traditional robotics, achieving high dexterity—the ability to manipulate objects with the grace of a human hand—has often been treated as a problem of increasing degrees of freedom (DoF). The logic is intuitive: more motors allow for more precise control over every joint. However, this approach leads to heavy, expensive, and computationally "expensive" systems that struggle to adapt to the messy, unstructured reality of the physical world.
A growing body of research suggests a different path: underactuation. By using fewer actuators than joints and linking those joints through clever mechanical "synergies," researchers are building robots that are simpler to control, more robust, and naturally adaptive.
Offloading Intelligence to Hardware
The core intuition behind this research is that functional dexterity can be achieved through hardware-embedded constraints rather than high-density active control. As Agneta Gustus and Georg Stillfried argue in Human hand modelling: kinematics, dynamics, applications, the human hand’s biological tendon structures provide a form of mechanical coupling. This coupling allows us to perform diverse grasping tasks without needing to think about every individual joint.
In robotic design, this translates to using a single motor to drive multiple joints. For example, the Hoecken-D Hand developed by Wenzeng Zhang uses a modified linkage and a spring-loaded differential to switch between precise pinching and adaptive enveloping. The "intelligence" of the grasp isn't in the software; it's in the way the mechanism naturally conforms to the object's shape.
Mechanical Synergies and Dimensionality Reduction
If a hand has 20 joints but only 5 motors, how does it decide how to move? The answer lies in synergies—the mathematical or mechanical coupling of degrees of freedom.
Zilin Si and colleagues demonstrate this in DELTAHANDS, where they use soft Delta robots as fingers. By reducing the control dimensionality through these synergies, they found that lower-dimensional control spaces can actually be more effective for manipulation than high-DoF systems. This is because the mechanical constraints automatically filter out "unnatural" or "unstable" configurations.
Similarly, the HIRO Hand presented by Dehao Wei simplifies the mapping between human movement and robotic execution by coupling distal and proximal finger joints. This reduces the mathematical complexity of the control system while maintaining the ability to replicate the vast majority of human grasping patterns.
Passive Adaptation through Tendons and Linkages
Underactuated systems often rely on tendon-driven mechanisms or complex linkages to distribute force. In the HI-ARM flying robot, Yuze Wu and Fei Gao use a single actuator to pull a flexible cable through a series of joints. This allows the "flying hand" to wrap around diverse objects passively, without requiring a dense array of sensors to tell the robot exactly where the object is.
This principle of "structural intelligence" extends beyond hands:
- Soft Robotics: Bobing Zhang describes an octopus-inspired robot that drives multiple soft arms through a single centralized crank-slider mechanism. This minimizes failure points while maintaining high maneuverability.
- Origami Metastructures: Yanbin Li and Hao Su 0002 have developed hierarchical origami that can transform into over 1,000 shapes using fewer than three active motors by nesting closed-loop mechanisms.
Why It Works: The Power of Constraints
The success of these systems stems from a fundamental shift in perspective: constraints are not limitations; they are features.
In a fully actuated system, the controller must solve a complex optimization problem to ensure all fingers coordinate to hold a cup. In an underactuated system like the Krysalis Hand, coupling the lateral movements of multiple fingers to a single gear train ensures coordination by design.
By prioritizing mechanical robustness and clever kinematics over raw motor count—as seen in the TriFinger platform by Manuel Wüthrich and Felix Widmaier—researchers are creating robots that can learn complex tasks like sliding and rotating objects through days of autonomous reinforcement learning without breaking.
Ultimately, this direction suggests that the future of robotics lies not in building more complex "brains," but in designing more "intelligent" bodies that simplify the world for the brains they have.