Skeleton Mamba: Efficient SSM-Based Skeleton Modeling

• Skeleton Mamba is a family of SSM-based architectures that integrate anatomical priors with multi-directional scanning to capture spatial, temporal, and spatio-temporal dependencies in 3D skeleton data.
• Key implementations such as SkelMamba and Parts-Mamba demonstrate robust performance in action recognition, pose lifting, and generative motion synthesis while minimizing computational overhead.
• Hybrid designs combining SSMs with GCNs, self-attention, and Transformers drive practical applications in clinical diagnostics, surveillance, and real-time biomechanical analysis.

Skeleton Mamba designates a family of state-space model (SSM)-based architectures that integrate structured, multi-directional scanning strategies to efficiently capture spatial, temporal, and spatio-temporal dependencies in skeletal data. These models are unified by the goal of anatomically faithful, computationally efficient, and long-range dependency modeling for 3D skeleton-based action recognition, pose estimation, medical diagnostics, and generative sequence modeling. Skeleton Mamba approaches draw heavily from the Mamba SSM paradigm, extending it to multidimensional skeleton sequences by partitioning joint and time dimensions and exploiting domain-specific anatomical priors. Key implementations include SkelMamba for clinical action recognition (Martinel et al., 2024), Parts-Mamba for occlusion-robust recognition (Shen et al., 21 Nov 2025), topology-aware variants for pose lifting (Zheng et al., 27 May 2025, Cui et al., 12 Nov 2025), and generative modules for motion synthesis (Tang et al., 9 Jul 2025, Nguyen et al., 14 Aug 2025).

1. Core Principles and Mathematical Foundations

The Skeleton Mamba design philosophy is anchored in the mathematical formulation of SSMs as the computational backbone for skeleton sequence modeling. The continuous-time SSM is formalized as

$$\dot{\mathbf{h}}(t) = A\,\mathbf{h}(t) + B\,x(t), \qquad y(t) = C\,\mathbf{h}(t)$$

where $A$, $B$, $C$ are learnable parameters and $x(t)$ denotes the time-indexed input (opportunistically embedding joint coordinates or part-grouped features). This is discretized via zero-order hold (ZOH) to permit efficient scan or convolutional processing: $$\mathbf{h}_t = \overline{A} \,\mathbf{h}_{t-1} + \overline{B}\, x_t, \qquad y_t = C\,\mathbf{h}_t$$ Dynamic parameterizations (as in the Mamba class of models) allow $A$, $B$, $C$ (and the discretization interval) to be functions of the input, supporting input-adaptive evolution and selective long-range integration. Channel-partitioning enables multiple independent or group-specific SSMs to operate in parallel, each specialized for a part, modality, or scan direction.

The core technical innovation enabling Skeleton Mamba models to scale to spatial-temporal skeleton data is multi-directional scanning. Canonical examples partition the input tensor $X \in \mathbb{R}^{T \times V \times C}$ (frames $A$0, joints $A$1, channels $A$2) along channel and/or joint dimensions and apply SSMs in mutually orthogonal or bidirectional axes (temporal-to-spatial, spatial-to-temporal, and reverses), capturing both local interactions and global dependencies at low complexity (Martinel et al., 2024, Liu et al., 1 Jun 2025).

2. Anatomical Decomposition and Structural Priors

A central design element in Skeleton Mamba approaches is the explicit incorporation of anatomical structure. In SkelMamba (Martinel et al., 2024), the spatio-temporal stream is decomposed across $A$3 anatomical partitions (e.g., legs, torso, arms, inter-limb coordination), each processed via a dedicated part-specific SSM. The outputs of part SSMs, concatenated with a global SSM operating on the entire skeleton, are aggregated via a learned weighted sum: $A$4 Similarly, Parts-Mamba partitions the joint set into $A$5 (potentially overlapping) groups and applies part-level SSMs and a body-wide SSM in parallel, followed by topological (graph-based) aggregation and gated fusion. These mechanisms enable robust encoding of both local (part-specific) and global (whole-body) motion cues, which is critical for applications such as clinical gait analysis in neurology, where impairments are often localized or display subtle cross-body patterns.

Structure-aware modules have independently been leveraged for skeleton topology preservation in pose lifting. SasMamba (Cui et al., 12 Nov 2025) introduces a structure-aware spatiotemporal convolution (SA-Conv) module that predicts local sampling offsets along the kinematic graph, followed by stride-based SSM scans for multi-scale spatial context aggregation. Such modules circumvent the loss of adjacency information that arises from purely flattening the skeleton into a 1D sequence, ensuring both topological faithfulness and global receptive fields.

3. Multi-Directional and Multi-Scale Spatio-Temporal Modeling

State-of-the-art Skeleton Mamba architectures universally apply multidirectional or multi-scale scanning across the spatiotemporal skeleton manifold. In SkelMamba (Martinel et al., 2024), channels are subdivided and assigned to SSM blocks operating in each of four canonical directions: $A$6 where each SSM block processes $A$7 channels and captures unique directional dependencies.

TSkel-Mamba (Liu et al., 12 Dec 2025) models temporal dependencies with bidirectional Mamba SSM blocks per joint, integrating multi-scale temporal interaction (MTI) modules. These modules interleave standard temporal SSMs with cycle-wise channel aggregation operators over several kernel sizes, thus capturing both immediate and extended temporal context and enforcing cross-channel mixing beyond per-channel independence.

SasMamba leverages stride-based sampling, splitting features into streams with strides $A$8 along the joint axis, thereby aggregating information at multiple spatial resolutions. The outputs of separate bidirectional SSMs (temporal and spatial) are fused to form the block output.

4. Hybridization with GCNs, Attention, and Transformers

Several Skeleton Mamba variants employ hybrid designs that combine graph convolutional networks (GCNs), self-attention, and SSMs to leverage their complementary strengths. Parts-Mamba (Shen et al., 21 Nov 2025) and Simba (Chaudhuri et al., 2024) embed SSMs within GCN backbones, utilizing the GCN head for initial feature extraction and structure-aware message passing, followed by SSMs for long-range or part-specific dependency modeling.

TSkel-Mamba (Liu et al., 12 Dec 2025) implements a hybrid Transformer–Mamba framework, stacking layers with a spatial transformer block (joint-wise self-attention parameterized by shortest-path distances on the skeleton graph) followed by a temporal Mamba block for cross-frame dynamics.

Mamba-Driven Topology Fusion (Zheng et al., 27 May 2025) injects GCN modules between the convolution and SSM stages (“GEM” block), producing enhanced locality-aware representations prior to temporal SSM modeling. This design ensures that global, linear-time SSM dependency modeling does not erode fine-grained topological cues.

5. Empirical Performance and Benchmarking

Skeleton Mamba architectures consistently achieve or surpass state-of-the-art performance on a range of public and clinical skeleton benchmarks, often with sharply reduced computational cost relative to transformer or traditional GCN models.

Benchmarking Summary

Method	Dataset/Task	Accuracy / MPJPE / Dice	Model Size / Compute	Key Finding
SkelMamba	NTU-RGB+D 60 (X-Sub/jbm)	93.4%	6.84M params/9.7G FLOPs	+3.2% vs transformers, 7.06ms/sample (≤0.5× latency)
SkelMamba	ND (4-class neurology)	99.64% (joints+bones+motions)		Best ever neurology gait analysis
Parts-Mamba	NTU-60 (part occlusion, mean)	84.4%	~3.3 GFLOPs/block	+4.1% > prior SOTA, >1000 FPS, robust to occlusion
SasMamba	Human3.6M (MPJPE/P1)	41.48mm	0.64M params/1.3G MACs	Outperforms 5× larger transformer-hybrids
TSkel-Mamba	NTU-RGB+D 60 (X-Sub, joints)	91.4% (joint), 93.1% (multi-strm)	2.4M params/8.2G FLOPs	SOTA accuracy, lowest inference latency
SpineMamba	CTSpine1K (Dice)	94.40%		+1.88% over nnU-Net, best known spinal segmentation

A plausible implication is that Skeleton Mamba’s joint exploitation of anatomical priors and efficient SSM-based recurrences yields incremental gains on hard recognition tasks and outlier robustness (e.g., occlusions, missing frames), particularly with limited computation (Martinel et al., 2024, Shen et al., 21 Nov 2025, Cui et al., 12 Nov 2025, Liu et al., 1 Jun 2025, Liu et al., 12 Dec 2025).

6. Specialized Applications and Generative Modeling

The explicit skeleton-oriented design and multi-directional scanning in Skeleton Mamba variants has driven impact beyond classification or regression. For instance, EgoMusic-driven dance motion estimation (Nguyen et al., 14 Aug 2025) and music-guided video synthesis (Tang et al., 9 Jul 2025) employ Skeleton Mamba/U-Net diffusion backbones to generate physically-plausible, temporally consistent human motion or video sequences. These models apply structured SSM blocks at varying scales (group scans, joint scans, temporal scans), with strong theoretical guarantees of equivariant function approximation over skeleton graphs.

SpineMamba (Zhang et al., 2024) extends the architecture to volumetric segmentation, using a residual visual Mamba block to couple local 3D convolutions with global SSM modeling, and shape-prior modules (VSP) to preserve vertebral identity.

Clinical translation is pronounced: SkelMamba’s anatomically-aware partitioning and efficient inference profile enable privacy-preserving, real-time motion analytics suitable for outpatient, telemedicine, and at-home rehabilitation use cases.

7. Limitations, Future Directions, and Practical Considerations

Despite strong empirical performance, Skeleton Mamba models retain structural constraints that motivate ongoing research. Fixed or hand-coded groupings may under-express cross-group dependencies unless a large block stack is employed (Liu et al., 1 Jun 2025, Shen et al., 21 Nov 2025). Linear-core SSMs can falter with highly non-stationary or strongly nonlinear motion unless fused with GCN/attention (Liu et al., 12 Dec 2025, Shen et al., 21 Nov 2025). Most benchmarks and validation use single-actor, non-occluded skeletons; performance under multi-person, in-the-wild, or partially-observed streams remains an open area (Liu et al., 1 Jun 2025, Shen et al., 21 Nov 2025).

Foreseeable directions include hybrid blocks with learned grouping strategies, self-supervised or generative pretraining (e.g., masked skeleton modeling), adaptation to multi-person or noisy skeleton sources, and integration of SSM principles into high-resolution vision backbones. The light computational footprint and privacy advantages position Skeleton Mamba as an enabling technology for edge-based, real-time, and regulatory-compliant biomechanical analysis across clinical, athletic, and surveillance environments.