MANO: Hand Model, NFV Management, & ML Variants

• MANO is a polysemous term with distinct meanings in 3D hand modeling, NFV orchestration, and machine learning, defined by context-dependent methodologies.
• In 3D vision, MANO utilizes linear blend skinning with learned deformations from high-resolution 3D scans to accurately capture human hand motion.
• In NFV and 5G, MANO orchestrates virtual network functions via NFVO, VNFM, and VIM, while ML variants apply novel optimization and attention mechanisms.

MANO is a polysemous technical term whose meaning depends on disciplinary context. In 3D vision and graphics, MANO denotes the learned parametric hand model “hand Model with Articulated and Non-rigid deformations,” introduced with SMPL+H for joint hand-and-body capture; in ETSI NFV and 5G systems, MANO denotes the Management and Orchestration framework centered on NFVO, VNFM, and VIM; and later work reuses the name for unrelated methods such as the Multipole Attention Neural Operator, a GUI agent, a manifold optimizer for LLM training, and Matrix Norm-based unsupervised accuracy estimation (Romero et al., 2022, Sciancalepore et al., 2018, Colagrande et al., 3 Jul 2025, Fu et al., 22 Sep 2025, Gu et al., 30 Jan 2026, Xie et al., 2024).

1. MANO as a parametric hand model

In hand modeling, MANO is a learned, skinned, statistical model of the human hand that combines articulated motion with non-rigid, pose-dependent deformations. Its canonical formulation follows the SMPL family’s linear blend skinning design: $$M(\beta, \theta) = W\!\big(T(\beta, \theta), J(\beta), \theta, \mathcal{W}\big), \qquad T(\beta, \theta) = \bar{T} + B_S(\beta) + B_P(\theta),$$ with shape-dependent joints obtained by a learned regressor (Romero et al., 2022). The original formulation also defines a low-dimensional synergy space,

$\theta = \bar{\theta} + Uz,$

where 6, 10, and 15 pose principal components explain approximately $81\%$, $90\%$, and $95\%$ of pose variance, respectively (Romero et al., 2022).

Contemporary implementations commonly describe MANO with shape coefficients $\beta \in \mathbb{R}^{10}$, pose parameters $\theta \in \mathbb{R}^{48}$, a triangulated mesh with $N=778$ vertices, and $K=16$ joints including the wrist (Wang et al., 2023). The per-vertex skinned deformation is written as

$$v_i(\beta, \theta) = \sum_{k=1}^{K} w_{ik} G_k(\theta, J(\beta)) \begin{bmatrix} B_{0,i} + (B_S\beta)_i + (B_P(\theta))_i, 1 \end{bmatrix}^{\top},$$

which makes explicit the coupling of blend shapes, the kinematic chain, and learned skinning weights (Wang et al., 2023).

The model was introduced to cope with low-resolution, occlusion, and noise in full-body 4D capture. It is learned from around 1000 high-resolution 3D scans of hands of 31 subjects in a wide variety of hand poses, and was designed to be realistic, low-dimensional, compatible with standard graphics packages, and able to fit any human hand (Romero et al., 2022). This combination of compactness and mesh fidelity is why MANO remains central in hand pose estimation, hand-object interaction, and avatar construction.

2. Extensions, datasets, and biomechanical variants of the hand-model MANO

Later work uses MANO both as a representation and as an annotation engine. HOGraspNet presents a real hand-object interaction dataset with full grasp taxonomies, 22 rigid objects from the YCB dataset and 8 compound objects, 99 participants aged 10 to 74, and 1.5M RGB-Depth of sparse frames with annotations. It provides labels for 3D hand and object meshes, 3D keypoints, contact maps, and grasp labels, and obtains accurate hand meshes by fitting MANO and HALO to multi-view RGBD frames (Cho et al., 2024). This positions MANO as part of the supervision pipeline for dense hand-object interaction data.

Several works extend MANO’s scope rather than merely applying it. DART augments MANO with a wrist-enhanced template mesh of 842 vertices and 1,666 faces, 50 daily 3D accessories, and 325 hand-crafted 2D texture maps at $\theta = \bar{\theta} + Uz,$0 resolution. Its DARTset contains 800K synthetic RGB hand images paired with MANO pose parameters, 2D keypoints, 3D joints, and full mesh labels (Gao et al., 2022). The extension is motivated by a limitation repeatedly noted in the literature: standard MANO provides geometry and articulation but disregards textures and accessories.

Biomechanical work modifies MANO more radically. MS-MANO integrates a musculoskeletal system with MANO, using Hill-type muscles and tendon geometry to drive the skeletal system through physiologically constrained torque trajectories. Its simulation-in-the-loop refinement method, BioPR, improves gSDF on DexYCB from MPJPE $\theta = \bar{\theta} + Uz,$1 to $\theta = \bar{\theta} + Uz,$2 and Deformer from $\theta = \bar{\theta} + Uz,$3 to $\theta = \bar{\theta} + Uz,$4, while adding approximately $\theta = \bar{\theta} + Uz,$5 ms of inference overhead (Xie et al., 2024). A plausible implication is that MANO’s geometric prior is often sufficient for reconstruction, but not for physiological realism.

Reconstruction methods likewise combine MANO with non-parametric predictions rather than replacing it. MeMaHand reconstructs two-hand meshes and MANO parameters simultaneously from a single RGB image using Mesh-Mano Interaction Blocks and a mesh alignment refinement module. On InterHand2.6M it reports MPJPE $\theta = \bar{\theta} + Uz,$6, MPVPE $\theta = \bar{\theta} + Uz,$7, AUC $\theta = \bar{\theta} + Uz,$8, and PROJ2D $\theta = \bar{\theta} + Uz,$9, outperforming IntagHand’s MPJPE $81\%$0, MPVPE $81\%$1, AUC $81\%$2, and PROJ2D $81\%$3; removing MANO tokens degrades MPJPE from $81\%$4 to $81\%$5 (Wang et al., 2023). The result is not that MANO replaces direct mesh prediction, but that parametric and non-parametric hand representations can be made complementary.

Generative modeling adopts MANO as a strong structural prior. HandDreamer uses MANO-based initialization, MANO silhouettes, and a hand skeleton guided diffusion process for zero-shot text-to-3D hand generation, together with a corrective hand shape guidance loss. On its reported test set it achieves CLIP L14 $81\%$6, FID $81\%$7, and HPSv2 $81\%$8 (Rosh et al., 6 Apr 2026). Here MANO functions less as the final output space than as a geometry prior that stabilizes score distillation.

3. MANO as Management and Orchestration in NFV and 5G

In NFV and 5G systems, MANO means Management and Orchestration. In the ETSI NFV architecture it comprises three main functional blocks: the NFV Orchestrator (NFVO), which performs service and resource orchestration; the VNF Manager (VNFM), which manages VNF lifecycle operations such as instantiate, configure, scale, migrate, update, and terminate; and the Virtualized Infrastructure Manager (VIM), which manages virtualized compute, storage, memory, and network resources within the NFVI (Sciancalepore et al., 2018). This is a foundational control-plane concept rather than a model.

The same architectural triad appears in 5G-oriented formulations where forwarding and compute are jointly optimized. Optimization-enhanced MANO for energy-efficient 5G networks describes NFVO as the central decision-maker that determines how many instances of each VNF to deploy, where to place them, and how to connect them so that end-to-end KPIs such as latency are satisfied; VNFM then executes lifecycle actions, while VIM controls the NFVI through systems such as OpenStack and OpenDaylight (Malandrino et al., 2019). In this literature, MANO is inseparable from multi-objective orchestration over compute, routing, energy, and latency.

Benchmarking work further formalizes MANO as an object of measurement. Proposed MANO-specific operational KPIs include On-boarding Process Delay, Deployment Process Delay, Run-time Orchestration Delay, and Quality of Decision, alongside functional KPIs such as resource footprint and supported VIM variety (Yilma et al., 2019). This framing treats MANO as an evaluable platform with performance characteristics that can be benchmarked independently of any particular VNF.

4. Autonomous, availability-aware, and slice-aware NFV MANO

A substantial literature studies automated or zero-touch MANO. z-TORCH addresses the joint optimization of orchestration and monitoring by clustering VNF KPI profiles into affinity groups and using reinforcement learning to adapt monitoring frequency and surveillance epoch length. It reports near-optimal QoD up to $81\%$9 under higher variance and monitoring load that stabilizes at approximately $90\%$0–$90\%$1 of the Instant Placement baseline (Sciancalepore et al., 2018). The central idea is that monitoring itself must be orchestrated.

OptiLoop embeds optimization within MANO for energy-efficient 5G. It jointly decides node activation, VNF placement, and routing using LP-relaxation-driven procedures called fixProblems and saveEnergy. On the reported real testbed, OptiLoop compute time averages approximately $90\%$2 s, while total network service instantiation averages $90\%$3 s, so orchestration actions dominate runtime; in large-scale emulation, OptiLoop achieves energy savings close to the Optimal baseline (Malandrino et al., 2019). This suggests MANO can be framed as an optimization loop whose computational overhead is often smaller than the actuation overhead of the platform.

Reinforcement-learning-driven MANO also appears in network slicing. KB5G proposes a knowledge-plane-based MANO framework for zero-touch slicing and introduces the twin-delayed double-Q soft Actor-Critic method, TDSAC, for joint resource and energy control. In its reported environment, TDSAC is evaluated with 3 slices, 20 APs, and up to 50 subscribers, and improves energy consumption and CPU utilization relative to TD3, DDPG, and SAC baselines (Rezazadeh et al., 2022). A related parameterized-action formulation, PAT, treats VNF resource control as a PAMDP and uses deep reinforcement learning to learn reconfiguration, deployment, and cloud offloading decisions (Roig et al., 2019).

Availability and mobility introduce further variants. HASFC provides a MANO-compliant framework for deploying service chains with an availability-cost trade-off; in the reported vIMS use case, one candidate chain reaches availability $90\%$4 at cost $90\%$5 (Mauro et al., 2021). A UTM-aware MANO framework for UAV services in MEC-NFV environments incorporates flight plans and predicted attachment points into an ILP placement scheme; runtime reaches up to about $90\%$6 seconds in the largest instances tested, and worst cases require on average approximately $90\%$7 replicates per VNF to satisfy latency targets (Bekkouche et al., 2022).

Open-source MANO deployments show the operational consequences of these abstractions. An OSM-WireGuard VPNaaS proof-of-concept reports service instantiation in 4 min 26 sec, reducible to 2 min 44 sec if WireGuard is pre-installed, and up to 5.3 times higher throughput with up to 41% lower latency than OpenVPN (Haga et al., 2020). Benchmarking OSM and ONAP finds that OSM-4 requires about $90\%$8 of the vCPU and $90\%$9 of the RAM required by ONAP-Beijing in the tested setup, while both platforms exhibit broadly similar aggregate deployment delays for the evaluated vCPE chain (Yilma et al., 2019).

5. Other methods that reuse the name MANO, Mano, or MaNo

Recent machine learning literature reuses the token for several unrelated methods. “MANO: Exploiting Matrix Norm for Unsupervised Accuracy Estimation Under Distribution Shifts” defines MaNo as a training-free estimator based on a data-dependent normalization of logits and an entrywise $95\%$0 norm of the normalized logit matrix. With $95\%$1 in experiments, it reports state-of-the-art performance across synthetic, natural, and subpopulation shift benchmarks (Xie et al., 2024). This usage is unrelated to both hand modeling and telecom orchestration.

“Linear Attention with Global Context: A Multipole Attention Mechanism for Vision and Physics” defines MANO as the Multipole Attention Neural Operator. It interprets attention as an interaction problem over grid points, computes near-field interactions exactly and far-field interactions via coarse “multipole” summaries, and achieves linear time and memory complexity with respect to the number of grid points. Reported results include MANO-tiny top-1 accuracy $95\%$2 on Tiny-ImageNet-202 and relative MSE $95\%$3 on $95\%$4 Darcy flow, outperforming the compared ViT and local-attention baselines in that setting (Colagrande et al., 3 Jul 2025).

“Mano Report” uses Mano as the name of a GUI agent built on UITARS-1.5-7B and trained with supervised fine-tuning, offline reinforcement learning, online reinforcement learning, and a verification module. Reported results include Mind2Web Cross-Task Element Accuracy $95\%$5, Operation F1 $95\%$6, and Step Success Rate $95\%$7, along with OSWorld-Verified average score $95\%$8 (Fu et al., 22 Sep 2025). The name here identifies an interactive agent, not an acronym expansion grounded in ETSI or hand geometry.

“Mano: Restriking Manifold Optimization for LLM Training” defines Mano as an optimizer that projects momentum onto the tangent space of a rotational Oblique manifold and alternates manifold orientation by $95\%$9. It maintains only first-moment momentum, reduces optimizer state relative to AdamW, and reports better final perplexities than AdamW and Muon, for example $\beta \in \mathbb{R}^{10}$0 on LLaMA-1.3B on Pile versus $\beta \in \mathbb{R}^{10}$1 for AdamW and $\beta \in \mathbb{R}^{10}$2 for Muon (Gu et al., 30 Jan 2026). The reuse of the name is entirely orthogonal to the older hand-model and NFV meanings.

6. Ambiguity, disambiguation, and scholarly usage

The supplied literature shows that MANO is not a stable cross-domain acronym. In hand-vision and graphics papers it denotes the parametric hand model; in telecom and NFV papers it denotes Management and Orchestration; in later machine learning work it denotes Matrix Norm, Multipole Attention Neural Operator, or a separate named system such as Mano (Romero et al., 2022, Sciancalepore et al., 2018, Xie et al., 2024, Colagrande et al., 3 Jul 2025, Fu et al., 22 Sep 2025).

This suggests that bare uses of “MANO” are often insufficiently specific in interdisciplinary writing. Case variation alone is not a reliable disambiguator, because MANO, MaNo, and Mano all appear as distinct research objects. In practice, precision usually comes from the expansion, the paper title, or the associated model equations: $\beta \in \mathbb{R}^{10}$3 signals the hand model (Romero et al., 2022), while the NFVO–VNFM–VIM triad signals the ETSI framework (Sciancalepore et al., 2018). For arXiv-oriented scholarship, citation by title or arXiv identifier is therefore more informative than the token alone.