Generating Personalized Lower-Limb Kinematics Across Walking Speeds Using Subject-Conditioned Diffusion

Technical Summary: Generating Personalized Lower-Limb Kinematics Across Walking Speeds Using Subject-Conditioned Diffusion

Problem Statement
Personalizing exoskeleton assistance requires user-specific gait data across diverse locomotor tasks, particularly varying walking speeds. However, acquiring this data typically demands repeated, costly, and time-intensive motion capture sessions, which are especially burdensome for clinical populations (e.g., stroke survivors). Existing generative models often focus on population-level gait representations rather than capturing user-specific characteristics, or they fail to effectively transfer subject-specific gait patterns across different locomotor contexts. The core challenge addressed is generating personalized lower-limb kinematics at unseen walking speeds using only a single observed speed, while preserving the subject's unique gait identity.

Methodology
The authors propose a subject-conditioned residual diffusion framework designed to synthesize gait trajectories for a target speed ($v_{tgt}$) given a source gait sequence ($x_{src}$) at a known speed ($v_{src}$).

• Residual Formulation: Instead of generating raw kinematic sequences from scratch, the model learns to predict the residual change ($\hat{r}$) required to transform the source trajectory into the target trajectory: $\hat{r}(x_{src}, x_{tgt}) = x_{tgt} - x_{src}$. The final target sequence is reconstructed as $\hat{x}_{tgt} = x_{src} + \hat{r}$. This approach aims to model subtle speed-dependent variations while preserving the underlying subject-specific gait signatures.
• Model Architecture: The framework utilizes a Transformer-based denoiser (based on the GaitDynamics backbone) conditioned on three inputs:
  1. Subject Gait: The source gait sequence ($x_{src}$) is encoded via a bidirectional LSTM into a subject embedding ($z_s$).
  2. Speed Parameters: A vector containing source speed, target speed, their difference, and their ratio is processed through an MLP to create a speed embedding ($z_{speed}$).
  3. Conditioning Mechanism: The subject and speed embeddings are combined and used to modulate the Transformer's hidden states via Feature-wise Linear Modulation (FiLM). This allows the model to shape its internal representation based on the specific subject and speed transition without adding these as direct input channels.
• Training Objective: The model is trained on able-bodied data using a variance-preserving diffusion objective. The loss function combines four terms:
  1. Weighted noise prediction loss (knee joints weighted twice as heavily as hip/ankle).
  2. Residual reconstruction loss.
  3. Identity preservation loss (encouraging near-zero residuals for identical speeds).
  4. Subject preservation loss (a contrastive loss ensuring generated sequences retain the source subject's embedding distinct from other subjects).
• Datasets: The model was trained on an able-bodied dataset (22 subjects, 0.5–1.85 m/s) and evaluated on held-out able-bodied subjects and an out-of-distribution stroke dataset (19 subjects, 0.3–1.2 m/s) without stroke-specific fine-tuning.

Key Contributions

1. Subject-Conditioned Pipeline: A novel architecture that combines a subject's source-speed gait with a desired target speed to condition a diffusion model, enabling identity-preserving generation.
2. Residual Generation Formulation: A strategy to synthesize trajectory changes rather than raw data, improving the modeling of speed-dependent variations while maintaining subject identity.
3. Evaluation of Identity Preservation: A comprehensive evaluation demonstrating that the model can synthesize biomechanically plausible gait across speeds from a single source sample, outperforming deterministic baselines.

Results

• Accuracy: On held-out able-bodied subjects, the model achieved a Mean Absolute Error (MAE) of 3.4° (Pearson's $r=0.95$). On out-of-distribution stroke subjects (without fine-tuning), it achieved an MAE of 6.0° ($r=0.86$).
• Personalization: The average personalization rank (where rank 1 indicates the generated gait is closest to the true subject compared to all other subjects) was 3.3 for able-bodied and 2.8 for stroke subjects, indicating strong preservation of individual gait identity.
• Baseline Comparison: The proposed diffusion model reduced the MAE by over 70% compared to supervised feed-forward baselines (MLP and TCN), which yielded errors of 12.25° and 11.43°, respectively.
• Data Efficiency: Using a single source speed yielded an accuracy within 0.4° of using four source speeds, suggesting that extensive data collection across multiple speeds is not strictly necessary for high-fidelity generation.
• Clinical Generalization: The model generalized to stroke gait patterns, preserving subject identity even for the paretic and non-paretic sides, with cadence errors remaining low (approx. 6–7% average percentage error).

Significance and Claims
The paper claims that subject-conditioned residual diffusion can synthesize personalized gait across speeds from minimal data (a single observed speed). This capability significantly reduces the data collection burden inherent in downstream exoskeleton personalization workflows. By generating user-specific data without requiring repeated motion capture sessions across all target speeds, the framework addresses a major bottleneck in deploying effective, personalized wearable robot interventions for both able-bodied and clinical populations. The work demonstrates that a generative model can capture speed-dependent gait changes and subject-specific identity simultaneously, offering a viable alternative to population-average models for exoskeleton control.