ISyHand: Low-Cost, Dexterous Robot Hand

• ISyHand is an open-source, low-cost robot hand featuring a 2-DoF articulated palm and anthropomorphic finger joints that enable robust in-hand manipulation.
• It is designed for ease of manufacturing and repair, being assembled in about four hours from off-the-shelf motors and 3D-printed parts.
• Performance benchmarks show that the articulated palm enhances dexterity and stability, achieving superior cube reorientation and grasping outcomes compared to fixed-palm variants.

ISyHand (pronounced “easy-hand”) is an open-source, low-cost, easy-to-manufacture, on-joint servo-driven dexterous robot hand whose defining feature is a 2-DoF articulated palm added to a four-digit hand to improve dexterity while preserving largely human-like finger kinematics (Richardson et al., 30 Sep 2025). The platform is positioned around a practical design goal: to provide a research hand that is inexpensive, assembled in about four hours, built from off-the-shelf motors and fasteners plus 3D-printed parts, and capable of meaningful dexterous in-hand manipulation, particularly cube reorientation trained with reinforcement learning in simulation and deployed on the real hand (Richardson et al., 30 Sep 2025).

1. Design rationale and conceptual position

ISyHand was introduced against a background in which high-performance robot hands often combine at least one of four constraints: high cost, closed-source or custom hardware, manufacturing and maintenance complexity, or non-anthropomorphic kinematics introduced to maximize dexterity at the expense of usability or teleoperation (Richardson et al., 30 Sep 2025). The hand is explicitly framed as a response to those constraints: a research platform that is cheap enough for broader adoption, quick to build and repair, modular and customizable, robust enough for repeated experimentation, and dexterous enough to matter in manipulation benchmarks (Richardson et al., 30 Sep 2025).

A central design tension in the paper is the tradeoff between anthropomorphism and dexterity. Anthropomorphic hands are treated as desirable because they can use standard human tools and support intuitive teleoperation, whereas many attempts to increase dexterity alter finger kinematics in ways that reduce anthropomorphic correspondence. ISyHand addresses that tension by keeping the finger joints anthropomorphic in structure and introducing additional dexterity through a 2-DoF articulated palm rather than through highly non-anthropomorphic finger mechanisms (Richardson et al., 30 Sep 2025).

The articulated palm is therefore not an auxiliary refinement but the main architectural claim. The paper repeatedly distinguishes ISyHand from rigid-palm hands such as Allegro and LEAP on this basis. In the authors’ interpretation, palm articulation allows the hand to cup objects better, supports tight grasping of smaller objects and wider spreading for larger objects, improves grasping and in-hand manipulation by changing palm geometry, and increases dexterity with only a modest sacrifice in anthropomorphism (Richardson et al., 30 Sep 2025). In real experiments, the articulated palm was also observed to provide stability and robustness by catching the cube on the outside hand edge and rolling it back toward the palm center (Richardson et al., 30 Sep 2025).

2. Mechanical architecture and kinematic organization

ISyHand is a four-digit robot hand with three fingers and a thumb, 18 total degrees of freedom, and one actuator per joint using on-joint Dynamixel servos (Richardson et al., 30 Sep 2025). The hand is approximately 255 mm long, 130 mm wide, 38 mm thick, and 620 g, and it is described as about 50% larger than the hand of an average adult male and similar in overall size to the Allegro and LEAP hands (Richardson et al., 30 Sep 2025).

The degree-of-freedom breakdown is explicit. The index, middle, and ring finger each have 3 DoF flexion and 1 DoF ab/adduction at the palm connection, for 4 DoF per finger. The thumb has 3 DoF flexion and 1 DoF thumb rotation at the palm connection, for 4 DoF. The palm contributes 2 articulated palm joints. The resulting total is

$3 \times 4 + 4 + 2 = 18 \text{ DoF}.$

The reduced variant ISyHand (flat) fixes the palm joints and has 16 DoF (Richardson et al., 30 Sep 2025).

The articulated palm is segmented so that portions of the hand base can move relative to one another rather than forming a single rigid plate (Richardson et al., 30 Sep 2025). This changes relative finger-base positions and orientations, so grasp geometry is not produced solely by finger joints. The paper does not present a formal analytical kinematic derivation in the sense of full forward kinematics, Jacobians, or dynamic equations; its treatment is design- and benchmark-oriented rather than analytical-model-oriented (Richardson et al., 30 Sep 2025). Numerical joint limits are also not given in the text, although simulation joint limits are taken from each hand’s URDF (Richardson et al., 30 Sep 2025).

The actuation strategy is fully on-joint and servo-driven. This is contrasted with tendon-driven and pneumatic systems on manufacturing, maintenance, and control grounds: each joint can be commanded directly to a desired position, and the servos provide high angular resolution and torque-limited position control (Richardson et al., 30 Sep 2025). The paper does not describe tendon transmissions, gear ratios, or coupling mechanisms, implying direct servo-driven joint actuation rather than tendon routing (Richardson et al., 30 Sep 2025).

3. Hardware implementation, materials, and manufacturability

The hand uses 12 × Dynamixel XL330-M288-T and 6 × Dynamixel XC330-M288-T motors (Richardson et al., 30 Sep 2025). The XL330 motors are used for flexion/extension—3 in each finger and 3 in the thumb—while the more powerful XC motors are used for finger ab/adduction, palm flexion, and thumb rotation (Richardson et al., 30 Sep 2025). The cost breakdown reported in the paper is 330 USD for the 12 XL330-M288-T motors and 620 USD for the 6 XC330-M288-T motors (Richardson et al., 30 Sep 2025).

Most non-motor components are 3D-printed. The main printed components use CPE HG100 filament for linkages and palm, costed at 35 USD, and Filaflex 60A for deformable fingertips, costed at 60 USD (Richardson et al., 30 Sep 2025). The design is stated to be robust in a variety of printable materials, including Onyx, PCTG, PLA, PLA+, and flexible TPU (Richardson et al., 30 Sep 2025). The fingertips are soft and deformable, printed in Filaflex 60A, and equipped with embedded fingernails for precision pinching (Richardson et al., 30 Sep 2025). The palm and finger surfaces are intentionally flat to support mounting of thin tactile sensors, and the fingertips can be modified to accommodate fingertip tactile sensors (Richardson et al., 30 Sep 2025).

A practical mechanical contribution is the cable-routing scheme. Motor cables are routed inside or along protected channels in the finger links: the first layer of the link routes cables, the second layer covers and protects them, and wires are pre-bent in the primary actuation direction to reduce wear (Richardson et al., 30 Sep 2025). The motors are daisy-chained along each link to a central hub on the back of the palm (Richardson et al., 30 Sep 2025). This is presented as a response to exposed finger wiring as a major weakness in existing on-joint servo hands such as LEAP (Richardson et al., 30 Sep 2025).

Electronics and control hardware include a Robotis U2D2 communication interface, a custom power injection board, and a 5V supply; the U2D2 is costed at 60 USD and the power-related components at 40 USD (Richardson et al., 30 Sep 2025). Control is via USB using Dynamixel Wizard or the Dynamixel SDK (Richardson et al., 30 Sep 2025). The total hardware and material cost is reported as approximately 1,300 USD, and the hand is said to be assembled in approximately four hours (Richardson et al., 30 Sep 2025).

Repairability is quantified directly. Five representative joints were disassembled and reassembled by an expert, yielding an average repair time of 5:08 min:s, with the longest repair at the outer palm joint taking 7:26 and the shortest at the index distal joint taking 3:22 (Richardson et al., 30 Sep 2025). This supports the paper’s characterization of ISyHand as a research-friendly platform with easily replaceable broken printed parts (Richardson et al., 30 Sep 2025).

4. Physical performance and benchmark methodology

The paper reports several direct physical evaluations. In a camera-based teleoperation test via MediaPipe, the hand opened and closed 10 times in 10.32 s, corresponding to an average cycle time of about 1 s per open-close cycle (Richardson et al., 30 Sep 2025). In a bucket-lifting power grasp experiment, testing started at 3 kg, increased by 0.5 kg per lift, was repeated up to 7 kg, and a final trial extended to 9 kg, with zero failures across all lifts (Richardson et al., 30 Sep 2025).

Pinching force was measured in three configurations, mean over five trials. Thumb plus index flexion only produced 3.5 N; adding the middle finger and palm yielded 5.6 N; and adding thumb rotation yielded 5.5 N (Richardson et al., 30 Sep 2025). The paper interprets these measurements as suggesting that the palm and multi-digit coordination improve effective pinching capacity (Richardson et al., 30 Sep 2025).

The principal dexterity benchmark is in-hand cube reorientation in simulation. The comparative study includes ISyHand, ISyHand (flat), Allegro hand, and LEAP hand (Richardson et al., 30 Sep 2025). Simulation uses IsaacGym and PPO, reusing the action space, observation space, reward shaping, and PPO hyperparameters from the IsaacGym Allegro-hand setup, with only dimensionality changing according to hand DoF (Richardson et al., 30 Sep 2025). The paper does not provide explicit vector dimensions or policy-network architecture details in the text (Richardson et al., 30 Sep 2025).

A methodological contribution is the grid-based workspace evaluation. The grid origin is defined directly above the center of the joint axis connecting the middle finger to the palm at the height of the palm surface (Richardson et al., 30 Sep 2025). The grid spans $x: -10$ cm to $14$ cm and $y: -14$ cm to $18$ cm with 2 cm spacing, yielding 221 grid cell policies per hand (Richardson et al., 30 Sep 2025). Each grid-cell policy is trained for 5,000 epochs, corresponding to around 655 million simulation steps, with average training time per policy of 3.66 hours on an NVIDIA A40 GPU (Richardson et al., 30 Sep 2025). Evaluation uses 100 episodes of 20 s each, and the full grid is run with three different seeds per hand (Richardson et al., 30 Sep 2025).

The paper motivates this design by observing that reorientation performance strongly depends on cube position relative to the hand. The grid analysis is intended to reveal the best manipulation region for each hand, how performance varies across the palmar workspace, and whether the articulated palm changes the usable manipulation region (Richardson et al., 30 Sep 2025). This suggests that ISyHand is not only a hardware proposal but also a morphology-sensitive benchmark study.

5. Evidence for the articulated palm

The strongest evidence for the articulated palm comes from the comparison between ISyHand and the fixed-palm ablation ISyHand (flat). After 5,000 training epochs, the best grid-cell average consecutive cube reorientations are 31.46 for ISyHand, 24.16 for Allegro, 22.27 for LEAP, and 22.25 for ISyHand (flat) (Richardson et al., 30 Sep 2025). Workspace-wide metrics reinforce the same pattern: $\Sigma s$ is 1638 for ISyHand, 1084 for ISyHand (flat), 901 for Allegro, and 923 for LEAP (Richardson et al., 30 Sep 2025).

Thresholded counts show the same effect more clearly. For $s \ge 1$, the counts are 87, 79, 78, and 75 for ISyHand, ISyHand (flat), Allegro, and LEAP respectively. For $s \ge 10$, they are 70, 59, 42, and 47. For $s \ge 20$, they are 42, 17, 10, and 6. For $s \ge 30$, ISyHand has 3 cells and all others have 0 (Richardson et al., 30 Sep 2025). The authors therefore argue that the articulated palm provides not only a better best case but a larger manipulability region (Richardson et al., 30 Sep 2025).

The paper also investigates long-training behavior. Because 5,000 epochs did not show convergence, one additional policy per hand was trained on that hand’s best-performing grid cell for 300,000 epochs (Richardson et al., 30 Sep 2025). The reported training curves indicate that ISyHand learns fastest in the first 25,000 epochs, reaches near plateau around reward $x: -10$0 by $x: -10$1 epochs, and appears to converge around 200,000 epochs (Richardson et al., 30 Sep 2025). By the end, ISyHand, Allegro, and LEAP all lie roughly in the 15,000–17,000 reward range, whereas ISyHand (flat) is lower, around 13,000–14,000 (Richardson et al., 30 Sep 2025).

Final evaluation after 300,000 epochs yields median successful consecutive cube rotations of 71 for LEAP, 70 for ISyHand, 69 for Allegro, and 60 for ISyHand (flat) (Richardson et al., 30 Sep 2025). The paper reports paired Wilcoxon signed-rank tests with all pairwise differences statistically significant at $x: -10$2 after Bonferroni correction (Richardson et al., 30 Sep 2025). The empirical interpretation in the paper is therefore twofold: the articulated palm improves sample efficiency and workspace coverage, while long-run asymptotic performance among the top three hands becomes similar, with the fixed-palm variant remaining worse (Richardson et al., 30 Sep 2025).

6. Real-world deployment, comparative context, and limitations

For real-world policy deployment, the paper switches from the standard IsaacGym setup to DeXtreme, arguing that the standard environment is unlikely to yield policies robust enough for reality (Richardson et al., 30 Sep 2025). The real system consists of three ROS2 nodes: object-pose tracker, robot hand driver, and policy node (Richardson et al., 30 Sep 2025). Object tracking uses FoundationPose with an Intel RealSense D405 RGB-D camera, running at about 31 Hz on an NVIDIA GeForce RTX 2080 Ti (Richardson et al., 30 Sep 2025). The hand driver runs at 100 Hz, and policy deployment via ONNX Runtime runs at 30 Hz to match the default simulation control rate (Richardson et al., 30 Sep 2025). The manipulated object is a 3D-printed cube of approximately 95 g and 6.5 cm edge length (Richardson et al., 30 Sep 2025).

Across 30 two-minute episodes, the reported real-world results are a mean of 6.1 cube reorientations, a maximum of 16, 3 drops, 7 stuck episodes, and 10 pose-tracking failures, with the text stating that the 10 pose-tracking failure episodes were discarded (Richardson et al., 30 Sep 2025). The authors report that the hand rarely drops the cube and that the most common failure source is pose tracking rather than hand inability (Richardson et al., 30 Sep 2025). They also note several real-world contact issues: because the hard plastic links and cube had very low friction, they added 3D-printed flexible covers on the finger links using Filaflex 70A and tough tape on the palm (Richardson et al., 30 Sep 2025). Additional observed failure modes included the cube sticking on the friction wrap and cube corners becoming stuck between fingers or finger links, with slightly rounding cube edges proposed as mitigation (Richardson et al., 30 Sep 2025).

Within the broader literature summarized in the provided sources, ISyHand occupies a distinct position. EyeSight Hand is a fully actuated 7-DoF humanoid hand with integrated vision-based tactile sensors and compliant quasi-direct-drive actuation, emphasizing tactile-rich whole-hand manipulation rather than a low-cost articulated-palm morphology (Romero et al., 2024). SyLink Hand is a synergy-inspired linkage-driven anthropomorphic hand with 19 joints driven by 11 actuators, 520 g mass, and manufacturing cost of approximately USD 400, emphasizing mechanically embedded synergies rather than one actuator per joint (Wu et al., 12 Jun 2026). In the comparison table reported in the SyLink paper, ISyHand is listed with 18 actuated DoFs, dimensions $x: -10$3 mm, mass ~620 g, cost ~USD 1.3k, fingertip force ~5.5 N, and ~9 kg passive payload, under the category of on-joint servo-driven hands (Wu et al., 12 Jun 2026). This suggests that ISyHand sits between high-actuation, joint-servo research hands and more aggressively synergy-embedded linkage hands.

The limitations reported or implied in the ISyHand paper are also specific. The work does not provide a detailed analytical kinematic or dynamic treatment, major comparative performance claims are made in simulation, policies trained in the approximate comparative simulation are not expected to transfer directly without performance loss, and real-world cube manipulation required added friction augmentation not present in the base hardware description (Richardson et al., 30 Sep 2025). The dominant real failure mode was external pose tracking, indicating dependence on an external vision system rather than onboard tactile or perceptual feedback (Richardson et al., 30 Sep 2025). Real-world performance, with an average of 6.1 reorientations per valid two-minute rollout, is substantially below the strongest simulated results (Richardson et al., 30 Sep 2025). A plausible implication is that the articulated palm is strongly supported as a morphology-level advantage, while the broader closed-loop manipulation stack remains bounded by perception and contact-modeling limitations rather than by hand morphology alone.