- The paper introduces a 27-DoF robot that transforms between a 20-DoF anthropomorphic hand and a 25-DoF desktop humanoid without hardware replacement, using shared actuators, sensors, and modular mechanisms.
- The paper combines teleoperation, diffusion policies, reinforcement learning, ZMP-based locomotion, and keyframe authoring, achieving 72% average grasping success across 10 objects and 0.0019 m simulated body-position error.
- The paper demonstrates an end-to-end workflow involving reconfiguration, detachment, walking, box pushing, docking, and bottle manipulation, while highlighting unverified sim-to-real locomotion, tethered operation, and limited quantitative evaluation as key limitations.
Overview
Handroid is a desktop-scale, dual-embodiment robot that reuses a single 27-DoF electromechanical body as either an anthropomorphic dexterous hand or a desktop humanoid (2607.16187). The platform measures 0.33 m in height and weighs 2.05 kg, and is motivated by a morphological observation: both the human hand and body can be abstracted as branching topologies of articulated kinematic chains extending from a compact central structure (palm-to-finger versus torso-to-limb). The paper's central claim is that morphology need not be a fixed property of a robot; instead, the same articulated modules—joint arrangements, actuation, sensing, and contact geometry—can be repurposed across embodiments, enabling one physical system to support both contact-rich manipulation and whole-body mobility.
Mechanical and electrical design
In the Dexterous-Hand embodiment, 20 actuated DoFs are distributed over five digits, each providing one abduction–adduction and three flexion–extension DoFs, approximating a standard 21-DoF human-hand kinematic model. In the Humanoid embodiment, 25 articulated DoFs form a 4-DoF head, two 4-DoF arms, two 6-DoF legs (12 lower-limb DoFs total), and a 1-DoF central hip joint. Two additional prismatic DoFs drive rack-and-pinion sliding mechanisms integrated into the torso module, translating the arm modules downward so they become the index- and little-finger modules during transition to the hand configuration. Reconfiguration therefore requires no hardware replacement—only controlled module repositioning.
The electrical system centers on a stacked 40 mm × 80 mm mainboard integrating an ESP32-S3 controller commanding all Dynamixel actuators over a TTL bus, Wi-Fi state streaming to a host, gamepad-based remote operation, onboard storage of motion primitives for standalone execution, and IMU sensing on the mainboard plus at the index/middle fingertips (which serve as feet in the humanoid embodiment). Power is supplied either by onboard batteries or USB Power Delivery up to 140 W, with an STM32 monitoring supply and battery temperature. Three Dynamixel actuator types are allocated by torque requirements: high-torque XM430-W210-T units at hip-pitch and ankle-pitch joints, dual-axis 2XC430-W250-T units at finger abduction/flexion and hip-roll/knee-pitch joints, and compact XC330-T288-T units elsewhere. An electromagnetic flange provides approximately 180 N of holding force for rapid docking to and detachment from a Franka Research 3 arm.
Unified control stack
A single control and learning framework spans both embodiments. For manipulation, Handroid uses Apple Vision Pro-based teleoperation with DexPilot-style retargeting (running above 20 Hz), trains an object-conditioned Diffusion Policy with PointNet++ point-cloud encoding for grasping, and trains a PPO policy in IsaacLab for in-hand cube reorientation with domain randomization over friction, actuator parameters, object mass/scale, and initial states.
For locomotion, the paper presents three complementary controllers. RL tracking control combines a ZMP planner based on the fixed-height linear inverted pendulum model with LQR preview control, Mink differential IK (12 iterations per frame, warm-started) to produce time-indexed references from either gait parameters or keyframe-authored motions, and a DeepMimic-style tracking reward over root and body position, orientation, linear, and angular velocity terms. RL velocity control is reference-free, conditioned on commanded planar CoM velocity and yaw rate with an asymmetric actor-critic design using privileged critic observations. Keyframe motion control provides a Viser-based editor whose interpolated trajectories can be streamed directly through the Dynamixel position interface or exported as RL tracking references—a shared authoring path between open-loop playback and closed-loop policy training.
Experimental results
The evaluation addresses three questions: whether the reconfigurable body retains fine-grained manipulation capability, supports stable humanoid locomotion, and extends the task space beyond either embodiment alone.
Dexterous grasping was evaluated on 10 objects with diverse geometries using 100 teleoperated demonstrations (10 per object), FoundationPose 6D pose estimation from RealSense L515 RGB-D input, and 512-point object clouds. Success rates ranged from 5/10 (sprayer) to 9/10 (chip tube, sheep figurine), yielding a 72% average success rate under randomized object poses distinct from those seen in demonstrations. This result indicates that modest demonstration budgets suffice for object-conditioned generalization on this platform, though the per-object variance suggests sensitivity to geometry class.
In-hand reorientation policies trained in simulation were deployed zero-shot at 30 Hz on hardware, qualitatively maintaining a cube in hand while following target orientations—evaluated only qualitatively, without reported success metrics.
Locomotion results are primarily simulation-based: the tracking policy achieves a joint-position error of 0.12 rad and body-position error of 0.0019 m against ZMP-generated references in MuJoCo, while the velocity policy tracks a commanded 0.20 m/s forward velocity with 0.052 m/s error. Real-world humanoid validation relies on keyframe-authored motions (walking, turning, squatting, push-ups, pull-ups, pick-and-place) executed via direct joint-position playback rather than learned policies, meaning real-world deployment of the RL locomotion controllers is not demonstrated in the paper.
Long-horizon dual-embodiment task: the most integrative experiment chains embodiment switching, detachment from the Franka via the electromagnetic flange, obstacle-avoidance walking, box pushing, lying down, re-docking, and dexterous bottle pick-and-place. This demonstrates that embodiment switching composes with locomotion and manipulation within a single workflow, though it is presented as a single qualitative demonstration rather than a statistically characterized success rate.
Limitations and open questions
The paper concedes several constraints. All real-world humanoid experiments use prerecorded keyframe playback; neither the tracking nor the velocity policy is validated on hardware, leaving sim-to-real transfer of learned locomotion unverified. Tethered operation currently limits mobile tasks, which the authors identify as a motivation for future wireless operation. The grasping evaluation uses scanned meshes and a fixed camera pipeline dependent on FoundationPose, and the in-hand reorientation and long-horizon results lack quantitative success rates. Open questions include whether cross-embodiment policy transfer—sharing representations or data between the hand and humanoid configurations of the same modules—yields measurable benefits over single-embodiment training, and whether fingertip tactile sensing (currently absent) would improve both grasp stability and foot-ground contact estimation given the shared finger/foot modules.
Conclusion
Handroid demonstrates that a single 27-DoF electromechanical body can serve as both a functional anthropomorphic hand and a walking desktop humanoid, with a unified control stack spanning teleoperation, imitation learning, reinforcement learning, ZMP-guided gait generation, and interactive motion authoring. Its quantitative contributions—a 72% real-world grasping success rate from 100 demonstrations and sub-centimeter simulated body-tracking error—are solid but bounded by the absence of real-world validation of learned locomotion policies. The platform's principal value lies in enabling controlled study of morphology reuse and cross-embodiment learning on reproducible, low-cost hardware.