Bidirectional GaitNet: Anatomy-to-Gait Modeling

• The paper introduces a bidirectional framework that uses a simulation-driven forward model and a conditional VAE-based inverse predictor to map anatomy to gait.
• It employs a full-body musculoskeletal model with 23 bones, 22 joints, and 304 musculotendon units to capture high-dimensional biomechanics.
• Empirical evaluations demonstrate accurate gait cycle reconstructions and effective probabilistic muscle condition inference, especially for pathological gait.

Searching arXiv for the named model and closely related gait papers to ground the article with current references. Bidirectional GaitNet is a simulation-driven generative model that learns the relationship between human anatomy and gait in both directions: a forward model predicts gait from anatomy and gait conditions, and a backward model infers plausible anatomical conditions from observed gait. It was introduced as a model of human gait and anatomical conditions built on a full-body, simulation-ready musculoskeletal model with 23 bones, 22 skeletal joints, and 304 Hill-type musculotendon units, and it is explicitly aimed at the high-dimensional, non-unique inverse problem of estimating internal physiological conditions from walking motion (Park et al., 2023).

1. Definition and scope

The term “Bidirectional GaitNet” refers specifically to the 2023 model that links anatomy and gait bidirectionally through a simulator-distilled forward predictor and a conditional variational autoencoder-based backward predictor (Park et al., 2023). In this formulation, the forward direction maps anatomical and gait conditions to gait, whereas the backward direction maps gait, together with known skeleton and gait conditions, to a distribution over muscle conditions.

The model is situated in biomechanics and clinical gait analysis rather than biometric identification. This distinction is important because the term “GaitNet” is also used elsewhere for RGB-based gait recognition. In particular, the earlier “GaitNet” disentangles appearance, canonical, and pose features from RGB imagery and uses a 3-layer LSTM for recognition, but it is not the same model and is not bidirectional in the anatomy-to-gait sense (Zhang et al., 2019). Likewise, a later pose-based recognition paper introduced a “Siamese biGRU-dualStack Neural Network” and explicitly noted that it was not literally a GaitNet model, using GaitNet only as a baseline (Progga et al., 2024).

A central motivation for Bidirectional GaitNet is that anatomy–gait relationships are not one-to-one. Different muscular impairments can produce similar visible gait patterns, and similar anatomy can still yield somewhat different walking motions. This motivates a probabilistic inverse model rather than a single deterministic estimate (Park et al., 2023).

2. Anatomy and gait representations

The paper defines anatomical condition as

$$C_\mathrm{anatomy} = (C_\mathrm{skeleton}, C_\mathrm{muscle}).$$

The skeleton component is written as

$$C_\mathrm{skeleton}=(c_\mathrm{head},c_\mathrm{trunk},c_1,\cdots,c_8,\tau_1, \tau_2),$$

where the variables include scaling factors of the head and trunk, lengths of four upper and lower limbs relative to a healthy reference model, and femoral torsional angles corresponding to femoral anteversion and femoral retroversion (Park et al., 2023).

Each musculotendon is parameterized by weakness and contracture. Weakness is the ratio of maximum isometric force relative to the reference model, and contracture is a scaling factor of muscle length relative to the reference model. Although the full model has 604 muscle-condition parameters across upper and lower body, the learning setup retains lower-body muscles and some hip-attached muscles because upper-body muscle conditions correlate weakly with gait. The reported implementation dimensions are $(C_\mathrm{gait}, C_\mathrm{skeleton})$ with 13 dimensions, $C_\mathrm{muscle}$ with 268 dimensions, and a total conditioning vector referred to elsewhere as around 280 dimensions; the paper notes a slight reporting inconsistency between 268, 280, and 281 implied by the component counts (Park et al., 2023).

A gait pattern is represented as a time series of full-body poses,

$M=(Q_{0}, Q_{\delta}, Q_{2\delta}, \cdots),$

with each pose

$$Q=(\mathbf{h}, \mathbf{v}, \mathbf{q}_0,\mathbf{q}_1, \mathbf{q}_2,\cdots),$$

where $\mathbf{h}$ is root or pelvis height from the ground, $\mathbf{v}$ is root linear velocity parallel to the ground plane, $\mathbf{q}_0$ is global root orientation, and $\mathbf{q}_i$ are joint rotations relative to parent joints. Rotations are represented using the first two columns of the rotation matrix. A complete gait sample includes two gait cycles, parameterized by phase $$C_\mathrm{skeleton}=(c_\mathrm{head},c_\mathrm{trunk},c_1,\cdots,c_8,\tau_1, \tau_2),$$0, with 60 samples over the two-cycle sequence: $$C_\mathrm{skeleton}=(c_\mathrm{head},c_\mathrm{trunk},c_1,\cdots,c_8,\tau_1, \tau_2),$$1 Stride and cadence are included as gait conditions (Park et al., 2023).

3. Forward and backward models

The framework comprises Forward GaitNet and Backward GaitNet. The forward model approximates anatomy-to-gait simulation, but the paper simplifies whole-sequence prediction to a per-phase predictor: $$C_\mathrm{skeleton}=(c_\mathrm{head},c_\mathrm{trunk},c_1,\cdots,c_8,\tau_1, \tau_2),$$2 This reduces network size, avoids dependence on the sample rate $$C_\mathrm{skeleton}=(c_\mathrm{head},c_\mathrm{trunk},c_1,\cdots,c_8,\tau_1, \tau_2),$$3, improves accuracy, and makes the regression problem easier. Forward GaitNet is trained by supervised regression on simulator-generated tuples $$C_\mathrm{skeleton}=(c_\mathrm{head},c_\mathrm{trunk},c_1,\cdots,c_8,\tau_1, \tau_2),$$4, effectively distilling a predictive gait simulator into a fast surrogate model (Park et al., 2023).

The backward direction is constructed as a conditional VAE whose decoder uses the frozen Forward GaitNet. The inverse problem is simplified from estimating all anatomy to estimating muscle conditions given observed gait and known skeleton and gait conditions: $$C_\mathrm{skeleton}=(c_\mathrm{head},c_\mathrm{trunk},c_1,\cdots,c_8,\tau_1, \tau_2),$$5 The reconstruction pathway is expressed as

$$C_\mathrm{skeleton}=(c_\mathrm{head},c_\mathrm{trunk},c_1,\cdots,c_8,\tau_1, \tau_2),$$6

After training, the encoder serves as Backward GaitNet and yields a distribution over plausible muscle conditions rather than a single point estimate (Park et al., 2023).

A key design choice is that the latent variable is placed in the middle of the backward model rather than directly over predicted muscle conditions. The stated reasons are functional muscle correlations and better handling of multi-modality. The approximate posterior is regularized toward a standard normal prior $$C_\mathrm{skeleton}=(c_\mathrm{head},c_\mathrm{trunk},c_1,\cdots,c_8,\tau_1, \tau_2),$$7, and the latent dimensionality used in experiments is 32. The backward loss is

$$C_\mathrm{skeleton}=(c_\mathrm{head},c_\mathrm{trunk},c_1,\cdots,c_8,\tau_1, \tau_2),$$8

The third term regularizes predicted muscle conditions toward the reference healthy condition, which is encoded by the value $$C_\mathrm{skeleton}=(c_\mathrm{head},c_\mathrm{trunk},c_1,\cdots,c_8,\tau_1, \tau_2),$$9 (Park et al., 2023).

4. Data generation, sampling strategy, and implementation

Training data are generated using Generative GaitNet, described as a state-of-the-art predictive gait simulator. The paper reports approximately 500 hours of walking motions, generated in about 10 hours on a cluster with 20 Intel Xeon 6242 CPUs. Because exhaustive sampling of the very high-dimensional condition space is infeasible, the paper compares uniform random sampling and grid-based sampling. Grid-based sampling emphasizes corner points of parameter ranges, exposing the model to extreme pathological cases; although the full corner set would be astronomically large, about 1.7 million tuples were reported as sufficient to train both networks (Park et al., 2023).

The authors also train multiple backward models rather than a single one. Three different models are used: one trained only for knee and ankle muscle conditions, and two trained for the entire muscle set with different loss weights. This is described as analogous to consulting multiple doctors for different diagnoses, and it functions as a practical mechanism for covering diverse pathological regimes (Park et al., 2023).

Implementation details are explicitly reported. The models are implemented in PyTorch. The encoder has layers $(C_\mathrm{gait}, C_\mathrm{skeleton})$0, the pre-decoder has $(C_\mathrm{gait}, C_\mathrm{skeleton})$1, and Forward GaitNet has $(C_\mathrm{gait}, C_\mathrm{skeleton})$2. Hidden and output activations are LeakyReLU/Linear for the encoder, LeakyReLU/Sigmoid for the pre-decoder, and ReLU/Linear for the forward network. The reported dimensions are $(C_\mathrm{gait}, C_\mathrm{skeleton})$3, $(C_\mathrm{gait}, C_\mathrm{skeleton})$4, $(C_\mathrm{gait}, C_\mathrm{skeleton})$5, and $(C_\mathrm{gait}, C_\mathrm{skeleton})$6. Both forward and backward models are trained with Adam at learning rate $(C_\mathrm{gait}, C_\mathrm{skeleton})$7, with batch sizes 65536 for Forward GaitNet and 2048 for Backward GaitNet. Training hardware is a desktop with Ryzen 3950, NVIDIA 2070, and 64GB RAM; training times are 2 hours for Forward GaitNet and 1.5 hours for Backward GaitNet (Park et al., 2023).

5. Empirical evaluation

Evaluation is performed on unseen simulated data and on real patient gait. For unseen simulated evaluation, the paper uses 51 unseen simulated data points, including 7 pathology-oriented examples covering classic gait types and additional randomly sampled examples from valid anatomy-condition ranges. The reported average joint-angle reconstruction error over two gait cycles is less than 8 degrees, and the best setting is Grid-Grid, meaning grid-based sampling for both forward and backward model training. Representative Grid-Grid errors include pelvis at 8.48°, femur right and left at 11.87° and 11.16°, tibia right and left at 5.79° and 5.49°, talus right and left at 14.77° and 14.55°, spine, torso, neck, and head around 6.9°–8.3°, and shoulders around 1.8°–1.9° (Park et al., 2023).

Direct error in muscle-condition space is treated as inadequate because the inverse is non-unique. Instead, the paper samples 1000 muscle-condition vectors from the predicted posterior, jointly embeds those samples with ground truth using UMAP, and visually inspects whether the true condition lies in a high-density region. The reported result is that the ground truth lies near the predicted sample cloud. For Trendelenburg or hip-drop gait, the model produces multiple plausible physiological explanations, such as contracture on one side and weakness on the other side, which is presented as qualitative evidence for a probabilistic inverse mapping (Park et al., 2023).

For real patient data, the model is trained only on simulated data and then evaluated for sim-to-real generalization. Because true muscle parameters are unavailable, validation uses physical examination data. The paper describes implementing the physical exam in simulation by computing joint torque due to passive muscle forces and treating range of motion as maximal when torque magnitude exceeds 20 Nm. In one patient example, gait asymmetry leads the system to predict weaker right calf muscles than left, and the resulting simulated range-of-motion asymmetry aligns with observed exam tendencies. The study is stated to have been approved by the Seoul National University Hospital IRB (Park et al., 2023).

6. Interpretation, limitations, and relation to adjacent research

Bidirectional GaitNet is best understood as an anatomy–gait generative model rather than a recognition model. This distinguishes it from biometric gait-recognition systems that use bidirectional temporal modeling for identity matching. A later pose-based recognition system, for example, combined MediaPipe-extracted 3D body landmarks, Procrustes alignment, and a Siamese dual-stacked bidirectional GRU metric-learning model, but that work explicitly stated that it was not actually GaitNet (Progga et al., 2024). Other adjacent models use bidirectional sequence processing for different purposes, including BGaitR-Net for occluded gait sequence reconstruction with a Bi-LSTM (Kumara et al., 2021), a dual-dataset CNN-LSTM framework with bidirectional LSTM layers for explainable gait abnormality detection (Agarwal et al., 19 Sep 2025), and a Transformer masked autoencoder that uses non-causal self-attention to generate a subject-specific “normative twin” from abnormal gait (Motta et al., 2 Apr 2026). These works are related in mechanism or application, but they do not redefine the specific meaning of Bidirectional GaitNet (Kumara et al., 2021, Agarwal et al., 19 Sep 2025, Motta et al., 2 Apr 2026).

The limitations reported for Bidirectional GaitNet are biomechanically specific. Joint errors remain higher for talus or ankle and femur joints; upper-body muscle inference is weak because upper-body motion during walking is less constrained by gait mechanics; and the anatomy model omits the nervous system, a flexible tendon model beyond the chosen musculotendon abstraction, and skin. The input currently requires mocap, which the paper identifies as expensive and cumbersome in practice. The stated future directions are monocular video via video-to-mocap pipelines, more complex full-body or facial musculoskeletal models, richer anatomical models with volumetric muscles, and improved network architectures or losses, possibly with Transformers, especially for joints with larger variation (Park et al., 2023).

A plausible implication is that the model’s principal contribution is not merely bidirectionality in a sequence-modeling sense, but a probabilistic bidirectional coupling between latent anatomical conditions and dynamically generated gait. In that sense, Bidirectional GaitNet occupies a distinct position: it connects musculoskeletal simulation, surrogate forward modeling, and inverse inference in a single framework, with the backward direction explicitly designed to reflect the many-to-one nature of gait pathology (Park et al., 2023).