Handroid: One Robot That Walks and Manipulates with Reconfigurable Body
Imagine a robot that needs to enter a room, walk to a table, and pick up a cup to pour water. With current technology, this typically requires two machines—or at least a mobile base, a robotic arm, and a dexterous hand. But what if the same body could both walk and transform into a five-fingered hand? Handroid turns this either/or choice into the past.
That's the vision behind Handroid, a desktop-scale robot developed by the University of North Carolina at Chapel Hill and Stanford University. Standing 0.33 meters tall and weighing 2.05 kilograms, it has 27 degrees of freedom. Its hallmark is the ability to switch between two morphologies: a small humanoid for walking and whole-body interaction, and a multi-fingered dexterous hand for fine manipulation.
The team drew inspiration from human anatomy. The human body and hand share a similar topological structure—both consist of a central region from which five multi-jointed branches extend. The thumb corresponds to the head, the index and little fingers to the arms, the middle and ring fingers to the legs, and the palm to the torso. Based on this insight, Handroid reuses the same 27-DOF mechatronic system for both forms, switching via two sliding tracks and rack-and-pinion drives—no disassembly or part replacement required.
In its hand form, 20 joint DOFs create a five-fingered structure close to a human hand, supporting grasping, multi-finger coordination, and in-hand manipulation. In the humanoid form, the five fingers become the head, arms, and legs, with 25 joint DOFs forming the body—the legs alone have 12 DOFs, enabling walking, turning, side-stepping, squatting, and posture transitions.
To validate the design beyond mere shape-shifting, the team conducted experiments from simulation to real robot. In hand mode, a custom teleoperation system built on Apple Vision Pro collected demonstration data to train diffusion policies, achieving an average 72% success rate in grasping 10 object categories. In humanoid mode, reinforcement learning policies support forward/backward motion, turning, side-stepping, squatting, and push-ups.
The most telling demonstration is a long-horizon task spanning both morphologies: the robot starts as a hand attached to a robotic arm, reconfigures into a humanoid, separates from the arm, moves along a path, and pushes a box into the arm's workspace. It then switches back to a hand, reattaches to the arm, and picks up a bottle to place in the box.
From a research perspective, Handroid transforms 'morphological reuse' from a concept into a functional hardware platform. It offers cross-morphology research tools for robot learning and raises a new question for hardware designers: should future robot hardware prioritize single-function optimality, or embrace unified, multi-morphology, multi-task platforms? When the same joint can be reassigned to new roles across tasks, the boundaries of robot capability may depend less on adding new components and more on how existing bodies are reorganized.