Gait Foundation Model Overview

• Gait foundation models are pretrained, general-purpose models for locomotor data that learn reusable representations and enable cross-domain transfer.
• They integrate diverse modalities such as IMU streams, silhouettes, and 3D skeletal sequences to support various applications including anomaly detection and clinical assessment.
• Innovative learning objectives like self-supervised masking and transformer-based encoding drive enhanced performance in tasks ranging from recognition to simulation.

A gait foundation model is a pretrained, general-purpose model for locomotor data that learns reusable representations or priors transferable across downstream tasks, subjects, conditions, sensing setups, or anatomical states. In the recent literature, the term encompasses several distinct but related formulations: self-supervised wearable encoders for inertial measurement unit (IMU) streams, large-scale silhouette and skeleton transformers, simulation-distilled generative models, and normative priors for anomaly detection and kinematic correction. What unifies these formulations is not a single architecture, but a common ambition: to replace narrowly task-specific gait pipelines with broad pretraining, reusable embeddings, and cross-domain transfer (Chi et al., 29 Sep 2025, Schwartz et al., 2024, Ye et al., 30 Nov 2025, Gabet et al., 26 Mar 2026, Park et al., 2023).

1. Historical lineage and conceptual scope

Classical gait analysis was organized around model-free appearance pipelines, model-based joint or link representations, and sensor-based measurements. The standard recognition pipeline emphasized preprocessing, temporal segmentation, representation, and matching; representative tools included Gait Energy Image (GEI), Dynamic Time Warping (DTW), Hidden Markov Models (HMMs), and silhouette similarity measures such as the Tanimoto score. These systems were designed chiefly for identification, verification, re-identification, and robustness to covariates such as viewpoint, clothing, carrying condition, surface, illumination, occlusion, and time lapse (Isaac et al., 2019).

A parallel biomechanical lineage treated gait as a structured dynamical system. Optimal-control formulations modeled movement trajectories as minimizers of costs involving force, yank, or metabolic surrogates, yielding conserved Hamiltonian-like quantities and closed-form walking trajectories under simplifying assumptions (Hagler, 2017). Other mathematical approaches focused on lower-limb inverse dynamics, joint torques, stride-level geometry, and constrained IMU-driven estimation in prosthetic and rehabilitation settings (JK et al., 2023). These approaches were “general” in a mechanistic sense, but they were not pretrained representation learners in the contemporary machine-learning sense.

Recent work uses “gait foundation model” more explicitly to denote a pretrained model learned on large-scale gait-related corpora that supports transfer across tasks, subjects, environments, or sensing configurations. In this usage, “foundation” may refer to a self-supervised encoder for wearable time series, a scalable silhouette representation learner, a 3D skeletal-motion encoder, or a bidirectional generative model linking anatomy and gait (Chi et al., 29 Sep 2025, Schwartz et al., 2024, Ye et al., 30 Nov 2025, Gabet et al., 26 Mar 2026, Park et al., 2023). This diversity suggests that the term now functions as an umbrella concept rather than a single standardized model class.

2. Modalities, corpora, and data regimes

Current gait foundation models are strongly shaped by modality. Wearable systems emphasize accelerometer and gyroscope streams; vision systems use silhouettes, 2D keypoints, or 3D skeletal motion; simulation-driven models encode musculoskeletal conditions; and generative clinical systems use pose or mesh trajectories. The scale and heterogeneity of the pretraining corpus are usually treated as constitutive features of the model family.

Representative model	Modality and pretraining source	Downstream emphasis
FM-FoG (Chi et al., 29 Sep 2025)	Diverse IMU gait corpora across PD and healthy cohorts	Real-time FoG detection and intervention
J-Net (Schwartz et al., 2024)	Wrist accelerometry encoder pretrained on UK Biobank	Gait segmentation amid HD chorea and PD
FoundationGait (Ye et al., 30 Nov 2025)	Binary silhouettes from 12 public datasets in WebGait-2M	Recognition, healthcare, and attribute estimation
Bidirectional GaitNet (Park et al., 2023)	Distilled predictive simulations of musculoskeletal gait	Forward synthesis and inverse anatomical inference
GAITGen (Adeli et al., 28 Mar 2025)	3D motion sequences from PD-GaM	Severity-conditioned gait generation
3D skeletal health model (Gabet et al., 26 Mar 2026)	3D depth-camera skeletons from deeply phenotyped adults	Multi-system phenotype prediction
GenGait (Motta et al., 2 Apr 2026)	Normative markerless 3D gait sequences	Joint-level anomaly detection and correction

FM-FoG pretrains on five IMU datasets spanning multiple sensor placements, sampling rates, and populations, including Parkinson’s disease (PD) and healthy cohorts, and then fine-tunes on VCU FoG-IMU for Freezing-of-Gait (FoG) detection (Chi et al., 29 Sep 2025). J-Net uses a ResNet-V2 foundation encoder pretrained with self-supervised learning on approximately 700,000 person-days from UK Biobank wrist accelerometry, then fine-tunes it with disease-specific labels and a segmentation head (Schwartz et al., 2024). FoundationGait pretrains on WebGait-2M, which contains 2,358,547 sequences and 231,675,844 frames from 12 public datasets spanning recognition and healthcare tasks (Ye et al., 30 Nov 2025). The 3D skeletal health model is trained on 3,414 deeply phenotyped adults, 17,589 sequences, and 351 hours of motion across five motor tasks (Gabet et al., 26 Mar 2026). Bidirectional GaitNet learns from approximately 500 hours of predictive-gait simulations and 1.7M condition–gait tuples, while GenGait is trained only on normative gait from 150 adults (Park et al., 2023, Motta et al., 2 Apr 2026).

A central implication is that “foundation” in gait is inseparable from dataset breadth. Papers repeatedly tie transfer claims to diversity in sensor placement, viewpoint, pathology severity, task protocol, or anatomical variation rather than to model scale alone (Chi et al., 29 Sep 2025, Ye et al., 30 Nov 2025, Gabet et al., 26 Mar 2026).

3. Architectural families and learning objectives

One dominant family uses self-supervised masked reconstruction on wearable or skeletal sequences. FM-FoG employs a 6-block encoder transformer with 8 attention heads and hidden dimension 128, conditions tokenization on sensor placement through learned location embeddings, and pretrains with masked IMU reconstruction on 1.28 s windows with 30% masked positions (Chi et al., 29 Sep 2025). The 3D skeletal health model uses a Dual-Stream Spatiotemporal Transformer (DSTformer) with masked denoising autoencoding, heavy spatiotemporal masking, and a combined reconstruction objective comprising MPJPE, NMPJPE, and velocity error (Gabet et al., 26 Mar 2026). GenGait նույնպես trains a transformer masked autoencoder, but on normative gait only, then uses a two-pass inference strategy to localize inconsistent joints and reconstruct a subject-conditioned “normative twin” (Motta et al., 2 Apr 2026).

A second family uses pretrained encoders plus task-specific heads. J-Net exemplifies this pattern: a ResNet-V2 feature extractor pretrained on UK Biobank wrist accelerometry serves as the contracting path in a U-Net-inspired segmentation architecture, and sample-wise masked cross-entropy is used for frame-level gait/non-gait segmentation (Schwartz et al., 2024). GaitFormer similarly pretrains a transformer on DenseGait using noisy multi-task learning, combining supervised contrastive learning on tracklet identities with multi-label attribute prediction from skeleton motion alone (Cosma et al., 2023). Its representation is reusable across identification, demographic estimation, and motion-only attribute prediction.

A third family is explicitly generative. Bidirectional GaitNet distills a predictive gait simulator into a forward network and embeds that forward model as the decoder of a conditional variational autoencoder, so that the learned system supports both conditions-to-gait prediction and gait-to-condition inference under uncertainty (Park et al., 2023). GAITGen uses a Conditional Residual Vector Quantized Variational Autoencoder with disentangled motion and pathology latents, followed by Mask and Residual Transformers for conditioned sequence generation across Parkinsonian severity levels (Adeli et al., 28 Mar 2025). These models treat a gait foundation model not merely as a feature extractor, but as a structured generator of clinically plausible motion.

FoundationGait introduces a different scaling-oriented design. It uses a DeepGaitV2-style CNN backbone, an EMA teacher–student setup, an InfoNCE objective, and a part-aware student that processes horizontal body partitions with different granularities $p \in \{1,2,4,8\}$. The stated purpose is to preserve local part diversity so that larger backbones continue to improve rather than saturate or degrade (Ye et al., 30 Nov 2025).

4. Downstream tasks and empirical performance

The downstream scope of gait foundation models is now unusually broad. In healthcare wearables, FM-FoG performs unseen-patient FoG detection and smartphone-triggered vibrotactile intervention. On the VCU FoG-IMU dataset, it reports 98.5% F1, 98.5% accuracy, 98.6% precision, and 98.5% recall across 25 random cross-subject splits, with total intervention latency of 17.5 ms on a Google Pixel 8a and battery-life improvement of up to 72% under event-triggered activation (Chi et al., 29 Sep 2025).

In neurodegenerative disease monitoring, J-Net addresses wrist-based gait segmentation under severe involuntary movement. It reaches ROC-AUC 0.97 for in-lab Huntington’s disease data, shows only a 10 percentage point drop from no chorea to severe chorea, achieves ROC-AUC $= 0.879 \pm 0.015$ in severe chorea versus $0.35 \pm 0.17$ for the baseline, and improves PD daily-living PR-AUC to 0.63 versus 0.55 for the baseline after fine-tuning on PD (Schwartz et al., 2024).

In gait biometrics and broad transfer, FoundationGait reports 48.0% zero-shot rank-1 on Gait3D and 64.5% on OU-MVLP with its 0.13B model, while fine-tuned performance reaches 79.3% rank-1 and 74.0% mAP on Gait3D. It also transfers to healthcare tasks, including 97.3% accuracy and 91.7% F1 on Scoliosis1K for the 0.03B model, and 78.2% instance-level F1 with linear probing on RA-GAR (Ye et al., 30 Nov 2025). GaitFormer, using skeleton-only motion with noisy multi-task pretraining, reports 92.5% accuracy on CASIA-B and 85.33% on FVG without manually annotated pretraining data (Cosma et al., 2023).

In clinical representation learning from 3D motion, the health-focused gait foundation model predicts age with Pearson $r = 0.69$, BMI with $r = 0.90$, and visceral adipose tissue area with $r = 0.82$, and significantly predicts 1,980 of 3,210 phenotypic targets. After adjustment for age, BMI, VAT, and height, gait embeddings still improve prediction in all 18 body systems in males and 17 of 18 in females (Gabet et al., 26 Mar 2026).

Generative systems serve different downstream functions. GAITGen improves downstream Parkinsonian gait severity estimation when synthetic data are added, increasing F1 from 0.67 with no synthetic augmentation to 0.73 with GAITGen-generated sequences (Adeli et al., 28 Mar 2025). GenGait preserves normative kinematics within a $\pm 1.5^\circ$ equivalence margin and significantly reduces RMSE in simulated gait abnormalities, with very large or large rank-biserial effect sizes for pelvis and hip angles (Motta et al., 2 Apr 2026). Bidirectional GaitNet, in turn, is positioned for fast forward gait synthesis, inverse diagnosis support, and personalized musculoskeletal simulation rather than conventional classification benchmarks (Park et al., 2023).

5. Generalization, robustness, and common misconceptions

A recurring misconception is that a gait foundation model is simply a large gait network. Several papers argue against this. FoundationGait states that classic gait architectures historically failed to follow scaling laws, because larger models tended to homogenize local part features; its part-aware student is proposed precisely to counter that problem (Ye et al., 30 Nov 2025). FM-FoG likewise argues that general-purpose time-series foundation models pretrained on non-physiological signals lack the inductive bias required for IMU gait, and empirically outperforms much larger time-series models with a compact 1.2M-parameter backbone (Chi et al., 29 Sep 2025).

Another misconception is that foundation-level transfer eliminates domain specificity. The literature instead emphasizes specialized inductive structure. FM-FoG uses sensor context embeddings and explicit resampling to 100 Hz to handle heterogeneous IMU configurations; a resampling ablation improves F1 from 91.2% to 98.5% (Chi et al., 29 Sep 2025). J-Net is robust to severe chorea not merely because of pretraining, but because sample-wise segmentation, skip connections, and triple-window context separate involuntary wrist motion from true gait (Schwartz et al., 2024). The 3D health model gains from late fusion across five motor tasks, and its anatomical ablations show that legs dominate metabolic and frailty predictions while torso carries sleep and lifestyle information (Gabet et al., 26 Mar 2026).

Deployment is also heterogeneous. FM-FoG is explicitly mobile and energy-aware: a lightweight CNN-LSTM gate activates the foundation model only during standing or walking, producing sub-20 ms end-to-end latency on-device and stable power traces (Chi et al., 29 Sep 2025). FoundationGait emphasizes zero-shot retrieval, lightweight fine-tuning of the last backbone blocks, and even transfer from silhouette pretraining to human parsing without modality-specific adapters (Ye et al., 30 Nov 2025). GenGait, by contrast, is not yet a broad deployment model; it is a label-free normative prior with interpretability at the joint level but limited scale, limited modality coverage, and no real patient cohort (Motta et al., 2 Apr 2026).

This suggests a more precise reading of the term. In gait research, “foundation” typically refers to reusable pretrained structure plus demonstrated transfer, but the transfer target may be identity, disease state, anatomy, intervention control, synthetic augmentation, or phenome-wide screening. The breadth of that transfer remains model-dependent rather than universal (Chi et al., 29 Sep 2025, Ye et al., 30 Nov 2025, Park et al., 2023, Gabet et al., 26 Mar 2026).

6. Limitations, controversies, and research directions

Despite rapid progress, current gait foundation models remain fragmented across modality, task, and validation regime. FM-FoG is evaluated on controlled clinical data rather than longitudinal in-home monitoring, and cross-device tests are not explicitly reported (Chi et al., 29 Sep 2025). J-Net uses wrist-only sensing, faces device heterogeneity across cohorts, and relies on daily-living face-validity rather than dense free-living ground truth (Schwartz et al., 2024). The 3D skeletal health model is trained on a cohort that is predominantly Ashkenazi Jewish and recorded with a single Azure Kinect setup, so ancestry fairness and sensor transfer are unresolved (Gabet et al., 26 Mar 2026). GenGait is evaluated on simulated abnormalities performed by healthy participants rather than on real pathology (Motta et al., 2 Apr 2026). GAITGen is specific to Parkinsonian gait and notes class imbalance in severe cases (Adeli et al., 28 Mar 2025).

Interpretability is another persistent issue. FM-FoG explicitly notes limited transformer interpretability (Chi et al., 29 Sep 2025). The health-focused 3D skeletal model provides anatomical-group ablations but not full mechanistic explanations of how joint coordination maps to molecular or organ-level phenotypes (Gabet et al., 26 Mar 2026). Bidirectional GaitNet partially addresses interpretability by linking latent variables to muscle weakness, contracture, skeleton scaling, and femoral torsion, but it inherits the domain gap and identifiability limits of simulation-only training (Park et al., 2023).

The main forward directions are comparatively consistent across papers. They include longitudinal and multi-site data collection, multimodal fusion across video, IMU, depth, audio, force, or EMG, stronger domain adaptation across sensors and environments, larger and more diverse pretraining corpora, longer temporal context, and better uncertainty estimation or explainability (Chi et al., 29 Sep 2025, Schwartz et al., 2024, Ye et al., 30 Nov 2025, Gabet et al., 26 Mar 2026, Motta et al., 2 Apr 2026). A plausible implication is that the field is converging on a layered view of gait foundation models: a reusable locomotor prior at the core, modality- and task-specific heads at the edge, and evaluation protocols that must span not only recognition benchmarks but also clinical generalization, robustness, fairness, latency, and privacy.

In that sense, the contemporary gait foundation model is less a single model family than a research program: to encode the fine-grained temporal, anatomical, and contextual structure of human locomotion once, then transfer it reliably across identification, pathology detection, anomaly localization, simulation, generation, and systemic health prediction (Ye et al., 30 Nov 2025, Chi et al., 29 Sep 2025, Gabet et al., 26 Mar 2026, Park et al., 2023).