Soft Polyhedral Network (SPN)

• SPN is a family of compliant, polyhedral-derived networks that distribute mechanical compliance via beams, joints, and lattice panels.
• It integrates embedded vision and proprioceptive learning to perform real-time 6D wrench estimation and enable adaptive in-hand manipulation.
• SPNs combine robust mechanical design with rigorous geometric frameworks, supporting omni-directional adaptation and multifunctional sensing in robotics.

Searching arXiv for the specified SPN papers to ground the article in current sources. Soft Polyhedral Network (SPN) denotes a family of soft robotic and geometric architectures whose structure is derived from polyhedra or polyhedral tilings, and whose compliance is distributed through beams, joints, lattice panels, or smoothly deformed edges. In robotics, the term has been used for a hollow polyhedral beam-network with embedded vision for proprioceptive learning and 6D wrench inference (Liu et al., 2023), for a lattice-like soft finger body that provides passive omni-directional deformation beneath an active tracked surface for in-hand manipulation (Li et al., 2023), and, in MagiClaw, for a vision-based proprioceptive fingertip in which an internal soft polyhedral-like lattice is observed by an embedded miniature camera and mapped to a 6-DoF contact wrench $\tau = [F_x, F_y, F_z, M_x, M_y, M_z]^T$ (Wu et al., 23 Sep 2025). In geometric terms, SPN also extends to smooth, symmetry-equivariant networks obtained by deforming polyhedral tilings while preserving combinatorics and space-filling structure (Domokos et al., 2024).

1. Definitions and terminological scope

The literature uses “Soft Polyhedral Network” in related but not identical senses. In "Proprioceptive Learning with Soft Polyhedral Networks" (Liu et al., 2023), an SPN is a class of soft robotic structures whose topology is derived from polyhedra: faces are removed and edges are replaced by compliant beams, then one or more internal layers are added and connected via flexure joints to form a hollow, networked metamaterial. In "Active Surface with Passive Omni-Directional Adaptation of Soft Polyhedral Fingers for In-Hand Manipulation" (Li et al., 2023), the SPN is a lattice-like finger body built from soft beams arranged on the edges of a polyhedral geometry with a pyramidal top. In MagiClaw, the SPN is a vision-based proprioceptive fingertip composed of a soft, polyhedral-like lattice observed by an embedded miniature camera (Wu et al., 23 Sep 2025). In "Soft cells, Kelvin's foam and the minimal surfaces of Schwarz" (Domokos et al., 2024), the concept is generalized to a smooth network obtained by deforming a combinatorial tiling while preserving adjacency structure and symmetry.

Source	SPN meaning	Core role
(Liu et al., 2023)	Hollow polyhedral beam-network with embedded vision	Proprioception, 6D force/torque inference, kinesthesia
(Li et al., 2023)	Polyhedral lattice finger body with pyramidal top	Passive omni-adaptation for tracked in-hand manipulation
(Wu et al., 23 Sep 2025)	Internal polyhedral-like lattice viewed by fingertip camera	Real-time 6-DoF contact wrench estimation
(Domokos et al., 2024)	Smooth network from deformed polyhedral tiling	Geometric foundation for soft, space-filling networks

Across these usages, the common invariant is a networked morphology derived from polyhedral structure and engineered to distribute compliance in three dimensions. The term therefore does not refer to a single fixed embodiment. A common simplification is to treat SPN as only a tactile sensor or only a soft finger morphology; the published record shows that it spans sensing, mechanics, manipulation, and geometric design (Liu et al., 2023).

2. Polyhedral topology, compliance distribution, and mechanical design

The SPN mechanical concept is based on replacing polyhedral edges with compliant beams or lattice members while preserving a structured, sparse, hollow geometry. In the proprioceptive SPN, the topology is a modified pyramid with two vertices on top, a trapezoidal primary interaction face for grasping, and a secondary face for spatial adaptation; outer edge beams and at least one inner layer are tied by flexure joints at mid-layer endpoints to reduce constraint coupling during large deformation (Liu et al., 2023). In the active-surface finger, all edges of a polyhedron are realized as soft-material beams, the top takes a pyramid shape with two vertices, the top beams are replaced with clevis pins to provide passive rolling guidance, and multiple layers form a lattice whose mid-layer edges terminate in curved joints to reduce interference during deformation while preserving support (Li et al., 2023).

This topology yields omni-directional passive adaptation. Reported deformation modes include conformal bending to curved objects, lateral and axial twisting around the central axis, and passive rolling guidance at the pyramid tip via clevis pins (Li et al., 2023). The proprioceptive SPN likewise emphasizes whole-body bending, twisting, and enclosure from multiple approach directions while maintaining a hollow interior for embedded vision (Liu et al., 2023). In MagiClaw, the fingertip lattice comprises repeated cell faces that behave like deformable panels; as the fingertip conforms to objects, the panels’ geometry distorts in characteristic, spatially structured ways, providing omni-directional conformability, structural integrity with localized deformation nodes, and high-density visually trackable features across the contact patch (Wu et al., 23 Sep 2025).

Material realizations differ by embodiment. The proprioceptive SPN is vacuum-molded as a monolithic soft structure from polyurethane elastomer, Hei-cast 8400, mix ratio 1:1:0, Shore 90A; calibrated FEM material properties are Young’s modulus $12.05\ \text{MPa}$, Poisson’s ratio $0.5$, and density $11.3\ \text{g/cm}^3$ (Liu et al., 2023). The active-surface SPN lattice is molded from a polyurethane elastomer, three-component mix ratio 1:1:0, hardness 90A, with a 3D-printed photosensitive-resin base frame that houses the timing belt and motor system; the reported SPN finger has weight $42\ \text{g}$, height $125\ \text{mm}$, and base width $54\ \text{mm}$ (Li et al., 2023). MagiClaw’s SPN fingertip is described as a flexible lattice, for example a TPU-based 3D-printed mesh, viewed by a miniature wide-angle micro-camera mounted inside the fingertip (Wu et al., 23 Sep 2025).

Stiffness is deliberately nonuniform. In the active-surface SPN, the fingertip segment is less compliant, which helps initiate contact and secure initial grasp, while the middle segment is more adaptive, promoting omni-directional conformal contact for manipulation; experimental characterization along the finger height $h$ and propulsion angles $\theta = 0^\circ, 45^\circ, 90^\circ$ shows anisotropic behavior, including a decrease from $1.699\ \text{N/mm}$ at $12.05\ \text{MPa}$0 to $12.05\ \text{MPa}$1 at $12.05\ \text{MPa}$2 for $12.05\ \text{MPa}$3, followed by a slight increase to $12.05\ \text{MPa}$4 at $12.05\ \text{MPa}$5 (Li et al., 2023). In the proprioceptive SPN, stiffness along the primary face is U-shaped, decreasing from approximately $12.05\ \text{MPa}$6 at H1 to approximately $12.05\ \text{MPa}$7 at H3 and slightly increasing to approximately $12.05\ \text{MPa}$8 at H4, while edges and the secondary face show decreasing stiffness with height (Liu et al., 2023). This suggests that SPN design uses anisotropy not as a defect but as a control variable for balancing secure initial capture against later conformal adaptation.

3. Embedded vision, proprioception, and learned wrench estimation

A major SPN lineage uses internal vision to infer deformation, force, torque, and, in some cases, shape. In the proprioceptive SPN, a miniature high-speed board camera with manually adjustable lens, $12.05\ \text{MPa}$9 FOV, and up to $0.5$0 is fixed in a base frame inside the structure and views an internal fiducial plate carrying a $0.5$1 ArUco marker at a marker-to-camera distance of approximately $0.5$2–$0.5$3 (Liu et al., 2023). The tracking pipeline performs ArUco detection and 6D pose estimation relative to an initial pose $0.5$4, producing the marker pose vector

$0.5$5

At $0.5$6, reported positional and angular noise ranges are approximately $0.5$7 and $0.5$8, the success detection rate is $0.5$9 over $11.3\ \text{g/cm}^3$0 consecutive frames in all tested resolutions, and pose computation time is approximately $11.3\ \text{g/cm}^3$1 per frame (Liu et al., 2023).

Force and torque estimation is learned from visual kinematic features. The proprioceptive SPN defines

$11.3\ \text{g/cm}^3$2

and uses features

$11.3\ \text{g/cm}^3$3

with velocity computed as a finite difference over 5 frames at $11.3\ \text{g/cm}^3$4 (Liu et al., 2023). The nonlinear regression model is

$11.3\ \text{g/cm}^3$5

Using deformation only, the reported MAEs are $11.3\ \text{g/cm}^3$6 in $11.3\ \text{g/cm}^3$7 and $11.3\ \text{g/cm}^3$8 in $11.3\ \text{g/cm}^3$9, with clear hysteresis; with deformation plus rate, MAEs improve to $42\ \text{g}$0 and $42\ \text{g}$1, and hysteresis is largely eliminated (Liu et al., 2023). Inference time is approximately $42\ \text{g}$2, with bandwidth limited by the camera to $42\ \text{g}$3.

The same paper distinguishes adaptive kinesthesia from viscoelastic proprioception. Adaptive kinesthesia maps marker pose to SPN shape $42\ \text{g}$4, represented as $42\ \text{g}$5 key points $42\ \text{g}$6 flattened to $42\ \text{g}$7 dimensions, using a Sim2Real MLP trained on $42\ \text{g}$8 FEM simulations; the reported mean positional error ranges from $42\ \text{g}$9 to less than $125\ \text{mm}$0, with average approximately $125\ \text{mm}$1 and inference approximately $125\ \text{mm}$2 (Liu et al., 2023). Viscoelastic proprioception introduces creep and relaxation modifiers. The relaxation modulus is modeled as

$125\ \text{mm}$3

with $125\ \text{mm}$4, $125\ \text{mm}$5, and $125\ \text{mm}$6; the creep compliance is

$125\ \text{mm}$7

with $125\ \text{mm}$8, $125\ \text{mm}$9, and $54\ \text{mm}$0 (Liu et al., 2023). These modifiers refine static predictions under sustained loading.

MagiClaw instantiates a related but distinct SPN sensing principle. Its embedded fingertip camera is modeled with standard pinhole intrinsics $54\ \text{mm}$1 and extrinsics $54\ \text{mm}$2, with projection

$54\ \text{mm}$3

Feature extraction includes corner and edge detection on lattice nodes and edges, optical flow between consecutive frames to obtain dense or semi-dense displacement fields $54\ \text{mm}$4, and a lightweight neural network computing

$54\ \text{mm}$5

under a supervised objective

$54\ \text{mm}$6

The paper emphasizes fingertip-specific calibration and training, real-time inference, and elimination of external force sensors during deployment, but it does not publish quantitative error metrics such as RMSE, resolution, bandwidth, hysteresis, or drift (Wu et al., 23 Sep 2025). This creates a clear distinction between the well-quantified proprioceptive SPN benchmark and the more systems-oriented MagiClaw deployment.

4. Manipulation architectures and task-level use

SPNs have been integrated into both passive-adaptive and actively driven manipulation systems. In the active-surface configuration, two SPN fingers augment a commercial DH-AG95 rigid gripper, and each finger carries one active surface comprising a TPU timing belt, a 3D-printed pulley with 2GT tooth profile, a Micro DC motor with pulley, and a driven roller (Li et al., 2023). The belt acts as a tank-like track. Synchronous belt motion, with $54\ \text{mm}$7, pulls an object upward from the fingertip into the more adaptive middle segment for repositioning; differential motion, with $54\ \text{mm}$8 and often opposite directions, creates a frictional couple that rotates the object for reorientation (Li et al., 2023).

The mechanics of this tracked manipulation are described by simplified balance equations. During repositioning,

$54\ \text{mm}$9

with standard friction model

$h$0

The maximum graspable weight at midpoint grasp is

$h$1

and, with $h$2 and motor torque $h$3, the reported value is $h$4 (Li et al., 2023). Reorientation follows

$h$5

and, under sticking and symmetric loading, a simplified pose Jacobian links belt speeds to vertical and rotational rates:

$h$6

Experimental results show repositioning success rates of $h$7 for a cube and $h$8 for an irregular vase, and reorientation success rates of $h$9 for a cylinder and $\theta = 0^\circ, 45^\circ, 90^\circ$0 for a sphere, with belts operated around approximately $\theta = 0^\circ, 45^\circ, 90^\circ$1 (Li et al., 2023). Failures in the vase trials typically occur when local convexity of SPN beams produces asymmetric contact, making one side move faster (Li et al., 2023). The paper attributes the system’s performance to a division of labor: the SPN provides passive, omni-adaptive substrate and stable conformal contact, while the active track supplies controllable traction for translation and rotation.

The proprioceptive SPN supports a different manipulation regime centered on sensing-rich control rather than belt-driven in-hand transport. Reported applications include sensitive and competitive grasping, in which two SPNs replace rigid fingertips on a commercial gripper, and touch-based geometry reconstruction, in which the SPN slides along an arch-shaped object while maintaining target contact force $\theta = 0^\circ, 45^\circ, 90^\circ$2–$\theta = 0^\circ, 45^\circ, 90^\circ$3 (Liu et al., 2023). In dynamic lift of a cylinder, the paper defines gripping force $\theta = 0^\circ, 45^\circ, 90^\circ$4, shear force $\theta = 0^\circ, 45^\circ, 90^\circ$5, and coefficient $\theta = 0^\circ, 45^\circ, 90^\circ$6 during sliding, with transition to static friction when $\theta = 0^\circ, 45^\circ, 90^\circ$7 exceeds the cylinder’s weight (Liu et al., 2023). Competitive grasping against a rigid gripper and a human finger is regulated in closed loop by maintaining $\theta = 0^\circ, 45^\circ, 90^\circ$8 within a friction cone, and disturbance rejection is demonstrated with four human pushes resisted with maximum $\theta = 0^\circ, 45^\circ, 90^\circ$9 (Liu et al., 2023).

MagiClaw integrates SPN fingertips into a dual-use, vision-based gripper architecture that is used both as a hand-held data collection tool and as a robotic end-effector (Wu et al., 23 Sep 2025). Its data pipeline streams fingertip images from miniature cameras to a Raspberry Pi, estimates fingertip wrench in real time, synchronizes $1.699\ \text{N/mm}$0, finger positions, and timestamps with iPhone ARKit streams of 6D pose, RGB, and LiDAR depth over Wi-Fi, visualizes data in an iOS app and Rerun, and exports synchronized records for offline imitation learning or reinforcement learning (Wu et al., 23 Sep 2025). Qualitative demonstrations include food handling and household tasks, including a multi-modal log for transferring a pizza slice to a plate and microwaving, and the system streams $1.699\ \text{N/mm}$1 back to the operator to provide haptic force feedback (Wu et al., 23 Sep 2025). The explicit systems claim is that preserving the same fingertip morphology and sensing during hand-held demonstration and robotic deployment reduces the human-to-robot domain gap (Wu et al., 23 Sep 2025).

5. Geometric foundations: soft tilings, symmetry, and equivalence classes

A separate SPN strand formalizes the geometry of soft, polyhedral-derived networks in terms of soft cells and soft tilings. A polyhedral tiling $1.699\ \text{N/mm}$2 in $1.699\ \text{N/mm}$3 is a space-filling partition with planar faces and straight edges, with no gaps or overlaps. A soft tiling $1.699\ \text{N/mm}$4 is a space-filling partition in which each cell boundary is everywhere at least $1.699\ \text{N/mm}$5, and no vertex of the combinatorial structure is realized as a sharp corner (Domokos et al., 2024). In this framework, an SPN is a smooth, symmetry-equivariant network obtained by deforming a polyhedral tiling’s 1-skeleton while preserving combinatorial structure and space filling (Domokos et al., 2024).

Two constructive procedures are identified. The universal edge-bending algorithm operates on a polyhedral tiling whose vertex duals have Hamiltonian circuits. At each vertex, all incident straight edges are simultaneously deformed into smooth curves so that their half-tangents become collinear; additional edge and face bending then removes sharp corners while preserving combinatorics, the node set, space filling, and symmetries (Domokos et al., 2024). In the Dirichlet–Voronoi tiling on the body-centered cubic lattice, this yields the standard soft cell $1.699\ \text{N/mm}$6. The bcc algorithm is a tiling-specific construction on the DV-bcc tiling. It begins with a planar quadrilateral face in $1.699\ \text{N/mm}$7, chooses a fundamental domain $1.699\ \text{N/mm}$8 as one-eighth of its $1.699\ \text{N/mm}$9-symmetric boundary, generates edges by reflection and by the bcc translation $12.05\ \text{MPa}$00, then completes the full $12.05\ \text{MPa}$01-edge cell by applying point-group operations $12.05\ \text{MPa}$02 (Domokos et al., 2024). Quadrilateral faces remain planar and hexagonal faces can be constructed as minimal surfaces consistent with boundary curves and symmetries.

The underlying combinatorics are fixed. For all members of the reported DV-bcc family, the cell is combinatorially equivalent to the truncated octahedron and has $12.05\ \text{MPa}$03 faces, $12.05\ \text{MPa}$04 vertices, and $12.05\ \text{MPa}$05 edges, with node valence $12.05\ \text{MPa}$06 and full symmetry group $12.05\ \text{MPa}$07 (Domokos et al., 2024). A one-parameter family $12.05\ \text{MPa}$08 connects the polyhedral truncated octahedron $12.05\ \text{MPa}$09, the standard soft cell $12.05\ \text{MPa}$10, the non-standard soft cell $12.05\ \text{MPa}$11, and Kelvin’s foam. If the fundamental domain $12.05\ \text{MPa}$12 is a circular arc with central angle $12.05\ \text{MPa}$13, then all edges become circular arcs with central angle $12.05\ \text{MPa}$14; it is often convenient to parameterize the family by $12.05\ \text{MPa}$15, where $12.05\ \text{MPa}$16 is the internal angle of a combinatorial hexagon and $12.05\ \text{MPa}$17 that of the quadrilateral, with node angular defect

$12.05\ \text{MPa}$18

Reported representative values include $12.05\ \text{MPa}$19 with $12.05\ \text{MPa}$20, $12.05\ \text{MPa}$21 with $12.05\ \text{MPa}$22, $12.05\ \text{MPa}$23 with $12.05\ \text{MPa}$24, and Kelvin’s foam with $12.05\ \text{MPa}$25, $12.05\ \text{MPa}$26, $12.05\ \text{MPa}$27, and $12.05\ \text{MPa}$28 (Domokos et al., 2024).

The paper further introduces two equivalence notions. First-order equivalence is combinatorial: two soft tilings are first-order equivalent if they share the same cell-face-edge-vertex incidence graph, node valence, and face-cycle structure. Second-order equivalence is defined by $12.05\ \text{MPa}$29 end-tangent data: within a first-order class, two soft tilings are second-order equivalent if, at each node, the ordered set of end-tangents of incident edges is identical up to rigid motion consistent with the full symmetry group (Domokos et al., 2024). Under DV-bcc symmetry and fixed nodes, exactly two second-order equivalence classes are reported: a standard class represented by $12.05\ \text{MPa}$30, with collinear half-tangents at nodes, and a non-standard class represented by $12.05\ \text{MPa}$31 and by the soft tiling induced by the approximate Schwarz P surface

$12.05\ \text{MPa}$32

whose Type-II section satisfies the required $12.05\ \text{MPa}$33 boundary smoothness condition (Domokos et al., 2024). In this sense, SPN design has both a robotic embodiment and a rigorous geometric classification.

6. Performance envelope, limitations, and open directions

SPNs combine low-cost compliant morphology with sensing or actuation, but the failure modes are specific to embodiment. In the proprioceptive SPN, performance is reported at $12.05\ \text{MPa}$34 with dynamic interaction accuracy of $12.05\ \text{MPa}$35 and $12.05\ \text{MPa}$36, durability beyond $12.05\ \text{MPa}$37 compression cycles at $12.05\ \text{MPa}$38 with $12.05\ \text{MPa}$39 stroke and approximately $12.05\ \text{MPa}$40 peak force, and impact capture up to $12.05\ \text{MPa}$41 over $12.05\ \text{MPa}$42; collision peak force is reduced from approximately $12.05\ \text{MPa}$43 for a rigid case to approximately $12.05\ \text{MPa}$44 for the soft case, with approximately $12.05\ \text{MPa}$45 safety improvement (Liu et al., 2023). At the same time, the paper identifies material aging and viscoelastic drift, larger $12.05\ \text{MPa}$46 MAE due to limited $12.05\ \text{MPa}$47-axis adaptation in the pyramid variant, and camera constraints imposed by off-the-shelf board hardware (Liu et al., 2023).

The active-surface SPN demonstrates that passive omni-adaptation alone does not eliminate grasp asymmetry. Failures arise when asymmetric contact due to local convexity causes uneven motion; traction loss occurs if $12.05\ \text{MPa}$48 per belt; hysteresis and damping in TPU belts and polyurethane lattice reduce control precision; and no tactile feedback or slip detection is used (Li et al., 2023). The paper proposes systematic scaling laws relating belt width, groove depth, contact area, load capacity, and motor torque; improved structural uniformity; tactile or force sensing; and extension to multi-finger SPN hands with coordinated belt control and richer Jacobian mappings (Li et al., 2023).

MagiClaw exposes a different set of systems-level constraints. Wireless latency and network congestion can bottleneck streaming and teleoperation; fingertip-specific calibration and training are required; reflective surfaces degrade LiDAR depth on the exteroceptive side; the iPhone is not hard real-time, so OS scheduling and thermal throttling can limit safety-critical control; and soft-material effects such as temperature sensitivity, material aging, and nonlinearity under high strain may require periodic recalibration, with saturation and dynamic effects under very fast impacts (Wu et al., 23 Sep 2025). The paper also notes that formal ablations and head-to-head comparisons to other tactile sensors are not reported (Wu et al., 23 Sep 2025). This suggests that MagiClaw’s contribution is primarily architectural and integrative rather than benchmark-driven.

A recurrent misconception is that SPN is equivalent to conventional vision-based tactile skins such as GelSight, DIGIT, or FingerVision. The proprioceptive and MagiClaw papers explicitly distinguish SPN from those systems: the deformable structure is an internal lattice or beam-network rather than an external gel surface, the sensor is fully embedded inside the soft fingertip or finger body, and the target is full 6-DoF wrench estimation or whole-body kinesthesia rather than only local surface geometry (Liu et al., 2023). Another misconception is that SPN is simply a variant of fin-ray fingers. The active-surface study states that fin-ray fingers predominantly flex in one plane and lose stiffness rapidly toward the fingertip, whereas the SPN tunes stiffness along height and orientation and adds twisting and rolling degrees of freedom via the polyhedral lattice and clevis pins (Li et al., 2023).

Future directions recur across the literature. Reported priorities include more sample-efficient calibration and pre-trained models for vision-based wrench estimation (Wu et al., 23 Sep 2025), more comprehensive state-space viscoelastic models and shared calibration across multi-SPN arrays (Liu et al., 2023), and integration of tactile or force sensing into active-surface SPNs for slip detection and feedback control (Li et al., 2023). On the geometric side, the continuous family linking standard soft cells, non-standard soft cells, the polyhedral DV-bcc cell, and Kelvin’s foam suggests a parameterized design space in which SPN connectivity, curvature, and node tangent patterns can be tuned while preserving combinatorics and symmetry (Domokos et al., 2024). A plausible implication is that future SPN systems may increasingly couple geometric class selection, compliance distribution, and learned sensing models as a single co-design problem rather than treating morphology, perception, and control as separate stages.