Handroid is a 27-degree-of-freedom desktop robot that reconfigures its own body from a fully articulated dexterous hand into a miniature humanoid. It uses the same motors, sensors, and control stack for both forms — no hardware swaps required.
What the Researchers Built
Handroid is a reconfigurable robot about the size of a desktop computer. It has six modular segments linked by Dynamixel servo actuators, giving it 27 degrees of freedom in total. Depending on how the modules are arranged, the same hardware becomes either a five-fingered dexterous hand or a torso with two arms, two legs, and a head.
The key innovation is a rack-and-pinion mechanism inside the main body that slides two of the modules up or down along a linear rail. Moving these modules from the "palm" position to the "torso" position completes the transformation in seconds. No tools, no replacement parts, no manual reassembly. The robot also carries a stacked mainboard (40×80 mm) with an ESP32-S3 controller, an STM32 power monitor, and onboard IMUs — one in the body and two in the fingertips/feet. It runs on batteries or USB-PD power, and streams state data over Wi-Fi for high-level policy control.
The control stack is shared across both embodiments. The same low-level actuator commands, proprioceptive feedback, and communication interface serve both dexterous manipulation and humanoid locomotion. This reuse means a policy trained for the hand can be transferred to the humanoid, and vice versa, without rewriting drivers.
Key Results
The researchers tested Handroid on three questions: can it match a dedicated dexterous hand, can it walk as a humanoid, and can it combine both embodiments in one task.
- Dexterous Hand embodiment: Using an Apple Vision Pro teleoperation setup to collect 100 demonstrations per task, Handroid achieved a 72% average grasp success rate across eight diverse objects (e.g., cup, screwdriver, apple, cube). It also performed in-hand cube reorientation — rolling a cube between its fingertips — and executed pick-and-place sequences.
- Humanoid embodiment: Handroid demonstrated stable walking using ZMP-guided reinforcement learning (RL) tracking control, RL velocity control, and keyframe-based motion playback. It walked, turned, and maintained balance on a flat surface.
- Long-horizon task: The robot executed a complete sequence that started as a humanoid, used its arms to interact with an object, then reconfigured into a dexterous hand, electromagnetically docked with a Franka Research 3 arm, and performed precise pick-and-place. This required no hardware changes between stages.
These results show that a single electromechanical platform can match the performance of purpose-built dexterous hands while also supporting humanoid locomotion.
How It Works
Handroid’s reconfiguration relies on two linear translation mechanisms built into Module VI (the central body segment). Each mechanism uses a small DC motor driving a rack-and-pinion gear set to slide a module along a rigid guide rail. During transition from humanoid to hand, Modules II and V move downward until they lock into the palm-finger configuration. The entire reconfiguration is controlled programmatically and takes about 20 seconds.
The electrical architecture is a vertically stacked mainboard (40 × 80 mm). The base PCB carries power regulation and an STM32 for battery monitoring and thermal protection. Above it, an ESP32-S3 board handles real-time control: sending position/velocity commands over a TTL bus to the 27 Dynamixel servos and reading back joint states at 100 Hz. Three IMUs (one body-mounted, two in the fingertips/feet) provide orientation and acceleration data. Wi-Fi streaming enables remote operation and policy deployment from a host PC, but the controller can also execute pre-stored motion primitives standalone.
For manipulation, the researchers built a VR teleoperation system using Apple Vision Pro. The operator’s hand motions are retargeted to Handroid’s hand, while wrist motion commands the end-effector of an external Franka arm. This setup collects demonstrations for imitation learning. For locomotion, they trained RL policies in simulation using ZMP (zero moment point) tracking as the reward signal, then transferred them to the real robot without retuning.
The control stack is shared: both embodiments use the same actuator IDs, sensor streams, and communication protocol. A policy written for the humanoid can directly command the hand’s joints in the right configuration, and vice versa.
Why This Matters for Robotics
Handroid challenges the assumption that dexterous hands and humanoid bodies need separate hardware platforms. If one robot can serve both roles, it reduces cost, simplifies development, and opens the door to cross-embodiment learning. A single robot could learn to manipulate objects as a hand and then walk across a room as a humanoid — all on the same physical system.
For the broader robotics industry, this points toward more flexible robot designs. Instead of buying separate grippers and bipeds, operations managers might eventually deploy reconfigurable platforms that adapt to different tasks on the fly. Handroid is still a research prototype, but its approach — shared modular actuators, a common control stack, and rapid reconfiguration — could influence future industrial robots, especially in R&D labs and small-batch manufacturing.
If you’re exploring humanoid platforms for research or demonstration, browse humanoid robots on Robot Overflow. For lighter manipulation tasks, check out used cobots for sale or used industrial robots.
Limitations and Open Questions
Handroid is still tethered during teleoperation, and cables can disturb balance in the humanoid embodiment. Wireless operation is a planned improvement. The actuators, while sufficient for desktop-scale tasks, limit the force output — heavier objects or higher walking speeds would require stronger motors. The lack of fingertip cameras and tactile sensors prevents fine in-hand pose estimation and makes foot-ground contact detection noisy. Finally, the current reconfiguration mechanism only works between two fixed morphologies; future work could explore continuous shape change. Scaling the design to human size would also face significant actuator miniaturisation challenges.
Frequently Asked Questions
How does Handroid reconfigure between hand and humanoid? Two modules slide along a linear rail driven by rack-and-pinion mechanisms inside the central body, repositioning fingers into legs in about 20 seconds.
Can Handroid walk and manipulate objects at the same time? In the humanoid embodiment it can walk and use its arms for whole-body manipulation, but precise dexterous tasks require the Dexterous Hand embodiment.
What was the grasp success rate? Handroid achieved a 72% average success rate across eight objects using imitation learning from just 100 demonstrations per object.
What control methods does Handroid support? It supports VR teleoperation via Apple Vision Pro, reinforcement learning policies (ZMP tracking and velocity control), keyframe playback, and standalone motion primitives.
Conclusion
Handroid proves that a single 27-DoF electromechanical system can serve as both a capable dexterous hand and a functional humanoid — without any hardware swap. Its 72% grasp success rate, stable locomotion, and successful long-horizon task demonstrate that reconfigurable platforms are viable for research in cross-embodiment manipulation and locomotion.