PaperlayerView exact revision on arXiv

Source-linked AI summary

Diffusion-Based Generation of Gait Trajectories

Damian Benasco, Juan Carballeira-Lopez, Jaime Ramos-Rojas, Julio S. Lora-Millan, Antonio J. Del-Ama, David Rodriguez-Cianca, Pablo Lanillos

arXiv:2609.14642v1cs.AIcs.LGcs.RO

TL;DR

Wearable-robot reference trajectories need to adapt to patient characteristics and goals while preserving biomechanical plausibility, a capability prior approaches do not provide through flexible continuous conditioning. The paper compares keypoint-conditioned transformer diffusion with a step-length-controllable Diffusion Transformer. The models show strong medium-step fidelity and shift short-step kinematics correctly, although short-step error is higher.

  • Problem

    Wearable-robot reference trajectories must adapt to patient morphology and therapeutic goals while maintaining biomechanical plausibility, but prior solutions lack flexibility for continuous, conditioned generation.

    Source: §I · p. 1

  • Method

    The paper compares a keypoint-conditioned transformer diffusion model with a Diffusion Transformer using adaptive normalization and classifier-free guidance for step-length conditioning.

    Source: §II · p. 1, C. DiT Variant: Controllable… · pp. 1–2

  • Results

    The base model reached R = 0.777 but converged to 57.4±3.7 cm; DiT achieved R = 0.838 for medium steps and R = 0.813 for short steps, with short-step NRMSE= 1.225.

    Source: §III · p. 2

  • Takeaways & Limitations

    The results support diffusion-based gait generation as a potential source of kinematic references for adaptive exoskeleton controllers.

    Source: §IV · p. 2

    Step length alone cannot distinguish clinically relevant gait patterns, since a slow shuffle and a fast short stride can share step length but differ in cadence and speed.

    Source: §IV · p. 2

Abstract

from arXiv · show

Generation of musculoskeletal gait trajectories conditioned on patient-specific parameters remains a key challenge for wearable robotics and rehabilitation. Assistive systems such as lower-limb exoskeletons require reference trajectories that adapt to individual morphology and therapeutic goals while preserving biomechanical realism. Traditional approaches rely on hand-crafted gait templates or optimization procedures that scale poorly across subjects and walking conditions. In this work, we explore conditional diffusion models for generating lower-limb joint-angle trajectories conditioned on gait parameters such as step length. We compare a baseline transformer diffusion model with a controllable diffusion transformer variant incorporating adaptive normalization and classifier-free guidance. Experiments on a dataset of 4,590 gait cycles show that diffusion models can generate realistic periodic gait trajectories while enabling some controllability variation in gait characteristics, highlighting their potential for personalized gait synthesis in assistive robotics.

I. INTRODUCTION

Wearable-robot reference trajectories need to adapt to patient characteristics and therapeutic goals while remaining biomechanically plausible. The paper investigates whether diffusion models can provide continuous, conditioned gait generation beyond the flexibility of prior approaches.

I. INTRODUCTION

  • Exoskeleton reference trajectories must adapt to patient morphology, pathology, and therapeutic goals while maintaining biomechanical plausibility.Prior approaches include hand-crafted templates and keypoint-based metaheuristics, but lack flexibility for continuous, conditioned generation.Source: §I · p. 1
  • Although diffusion models have shown results in motion synthesis, their application to clinical gait generation remains mostly unexplored.The passage also notes adaptive-control work using reinforcement learning for different morphologies.Source: §I · p. 1
  • The work evaluates keypoint-conditioned transformer diffusion and a controllable Diffusion Transformer incorporating physical gait parameters such as step length.The controllable variant is identified as a Diffusion Transformer.Source: §I · p. 1

II. METHODS

The methods frame gait synthesis as denoising diffusion and compare a keypoint-conditioned transformer with a controllable Diffusion Transformer extension. The models generate continuous lower-limb kinematic trajectories conditioned on biomechanical keypoints and, for the controllable model, step length.

II. METHODS

  • The framework pairs a baseline keypoint-conditioned Transformer diffusion model with a Diffusion Transformer extension for step-length controllability.The extension uses classifier-free guidance.Source: §II · p. 1
  • Gait synthesis is framed as a denoising diffusion process.This is the stated modeling framework for the compared models.Source: §II · p. 1
  • The diffusion model generates continuous lower-limb kinematic trajectories conditioned on biomechanical keypoints and step length via classifier-free guidance.The figure caption summarizes the generation setup.Source: Figure 1 · p. 1

A. Dataset

The dataset contains 4,590 gait cycles from 22 healthy subjects walking at short and medium step lengths. Each cycle is represented by 21 lower-limb joint angles and biomechanical keypoints, with a 4 cm exclusion gap separating the classes.

A. Dataset

  • 4,590 gait cycles from 22 healthy subjects were used, with 12 male and 10 female participants.Participants averaged 26.5±7.7 years of age, 172.1±7.9 cm in height, and 68.5±9.1 kg in weight.Source: A. Dataset · p. 1
  • Each gait cycle was represented by 21 lower-limb joint angles alongside biomechanical keypoints derived from VICON motion capture.Source: A. Dataset · p. 1
  • A 4 cm exclusion gap separated 1,756 short steps below 48 cm from 2,834 medium steps above 52 cm.Participants were instructed to walk at the two distinct step lengths.Source: A. Dataset · p. 1

B. Base Model: Keypoint-Conditioned Transformer

The base model treats each frame as a token and conditions on biomechanical keypoint signals by concatenating them across channels. It predicts the clean sample over a 500-timestep linear noise schedule.

B. Base Model: Keypoint-Conditioned Transformer

C. DiT Variant: Controllable Generation via AdaLN-Zero

The controllable DiT injects physical gait parameters through AdaLN-Zero at every layer and enables classifier-free guidance with a learned null embedding.

C. DiT Variant: Controllable Generation via AdaLN-Zero

  • AdaLN-Zero replaces layer normalization, using timestep and gait-parameter embeddings to modulate each layer’s shift, scale, and gate.Kinematic pulse conditioning remains channel-concatenated at the input.Source: C. DiT Variant: Controllable… · p. 1
  • Physical gait parameters are injected through AdaLN-Zero at every layer, while kinematic pulse conditioning remains channel-concatenated at the input.Source: C. DiT Variant: Controllable… · p. 1
  • Classifier-free guidance uses a learnable null embedding substituted during training with probability p=0.25.Source: C. DiT Variant: Controllable… · p. 2

D. Training Objective

The training objective combines reconstruction, periodicity, smoothness, and forward-kinematics foot-trajectory losses, with a curriculum for the forward-kinematics term.

D. Training Objective

E. Model comparison

The comparison evaluates waveform fidelity, biomechanical validity, and cross-subject generalization using correlation, error, range-of-motion, and symmetry measures.

E. Model comparison

III. RESULTS AND DISCUSSION

The conditioned DiT separates generated step-length classes, with stronger reported fidelity for medium steps than short steps. Short-step results also show greater ROM error and degraded symmetry.

III. RESULTS AND DISCUSSION

  • R = 0.838 and NRMSE = 0.920 for medium-step DiT; short-step conditioning yields 36.4±6.4 cm and R = 0.813 but NRMSE = 1.225.The base model has R = 0.777 and converges to a mean step length of 57.4±3.7 cm.Source: Table I · p. 2, §III · p. 2
  • The generated step-length distributions separate both conditioned classes from the base model’s unimodal output.Source: Figure 2 · p. 2, §III · p. 2
  • Class 1 symmetry is well preserved, whereas class 0 hip and knee symmetry indices are approximately 16–17%.Per-subject R and NRMSE are reported across 22 subjects.Source: §III · p. 2, Figure 3 · p. 2
  • Class 1 ROM error is within ±6◦, while class 0 systematically underestimates knee ROM by ∼10◦ and hip ROM by ∼6◦.Higher short-stride inter-subject variability (knee ROM CoV ≈42% vs. 26%) is linked to MSE-driven regression toward the mean.Source: Table II · p. 2, §III · p. 2

IV. CONCLUSIONS

The proposed DiT generates biomechanically plausible gait trajectories controllable by step length for use as kinematic references in adaptive exoskeleton controllers. Class 1 shows strong fidelity, while class 0 is more difficult; step length alone does not distinguish clinically relevant gait patterns.

IV. CONCLUSIONS

  • Class 1 achieves strong fidelity (R = 0.838), ROM error ≤±6◦, and well-preserved symmetry.The trajectories are intended as kinematic references for adaptive exoskeleton controllers.Source: §IV · p. 2
  • Class 0 is harder because distributional heterogeneity amplified by CFG null-embedding bias produces amplitude underestimation and degraded symmetry.Source: §IV · p. 2
  • Step length alone cannot distinguish clinically relevant gait patterns, motivating future conditioning on speed, subject-specific body dimensions, and multiple continuous parameters.The paper notes that a slow shuffle and a fast short stride can share similar step lengths while differing in cadence and speed.Source: §IV · p. 2
Loading 2609.14642v1…