Handroid: A Reconfigurable Robot That Is Both a Dexterous Hand and a Humanoid

In traditional robot design, humanoid robots and dexterous hands are typically separated: the former excels at mobility and whole-body interaction, while the latter relies on robotic arms and fixed workspaces. This division leads to system complexity, such as coordinate transformation and collision avoidance challenges when the robot's position changes under external cameras.

Handroid's core concept is morphology reuse: leveraging the topological similarity between the human body and the hand, the same 27-degree-of-freedom electromechanical system is reused in two forms. In the dexterous hand form, 20 degrees of freedom constitute a five-finger structure; in the humanoid form, 25 degrees of freedom constitute the head, arms, and legs, with 12 degrees of freedom in the legs supporting actions like walking, turning, side-stepping, and squatting.

This design does not add new components but reorganizes existing hardware, allowing the same set of modules to play different roles in different tasks, thereby integrating mobility and fine manipulation.

The morphology switch is accomplished through two sets of sliding rails and rack-and-pinion transmission mechanisms, converting actuator rotation into linear motion to move joint modules to new positions without disassembling hardware. The team also designed an electromagnetic flange for the Franka Research 3 robotic arm, providing approximately 180 N holding force for quick connection and detachment, enabling Handroid to move independently outside the robotic arm's workspace.

In the electrical system, the main control board measures approximately 40 mm × 80 mm, integrating actuator control, wireless communication, power management, temperature monitoring, and status feedback. Inertial sensors on the body and fingertips serve hand control and body motion perception in the two forms, respectively, achieving sensor reuse.

In the dexterous hand form, the team built an arm-hand collaborative teleoperation system based on Apple Vision Pro, tracking hand keypoints in real time and mapping them to Handroid, while mapping wrist motion to the robotic arm's end effector. Based on collected demonstration data, they trained an object-conditioned diffusion policy that takes object point clouds and proprioceptive states as input and outputs continuous action sequences.

The humanoid form supports two types of reinforcement learning approaches: tracking reference trajectories (e.g., from ZMP gait planners or keyframe-edited actions) and learning directly from target velocities. The team also optimized a keyframe editing tool based on Viser for rapid design and validation of actions.

In dexterous manipulation experiments, the team collected 100 demonstrations across 10 object types, and the trained policy achieved an average grasp success rate of 72% under random poses. In in-hand manipulation tasks, the reinforcement learning policy ran at 30 Hz, completing cube holding and reorientation.

In humanoid experiments, the reference trajectory-based policy achieved a joint position error of 0.12 radians and a body position error of 0.0019 meters in simulation; the direct velocity command policy achieved a velocity tracking error of 0.052 m/s under a forward command of 0.20 m/s. The real robot could perform forward/backward movement, turning, side-stepping, squatting, push-ups, and pull-ups.

In a cross-morphology long-horizon task, Handroid switched from the dexterous hand form to the humanoid form, detached from the robotic arm, moved and pushed a box, then switched back to the dexterous hand form to grasp a bottle and place it into the box, fully demonstrating morphology switching and task continuity.

Handroid's value lies in turning morphology reuse from a concept into a working platform, covering the full-stack development of mechanics, electronics, control, and learning. It provides a cross-morphology task carrier for robot learning research and proposes a new paradigm for hardware design that unifies multiple morphologies and tasks.

The project is a joint effort between the IDEAL Lab at the University of North Carolina at Chapel Hill and Stanford University, with guidance from Karen Liu, Shuran Song, and Mingyu Ding. The paper, CAD models, and BoM list are open-sourced on the project homepage.