Model-agnostic Meta-learning for Adaptive Gait Phase and Terrain Geometry Estimation with Wearable Soft Sensors

Zenan Zhu, Wei Chen, Pei-Chun Kao, Janelle P. Clark, Lily Behnke, Rebecca Kramer-Bottiglio, Holly A. Yanco, Yan GuIEEE RA-L 2026 · DOI

TL;DR

This paper introduces a novel framework using model-agnostic meta-learning (MAML) with wearable soft sensors to accurately estimate human gait phase and terrain geometry. This approach significantly improves adaptation to new users and terrains with minimal calibration data, outperforming traditional methods.

What Problem Does This Paper Address?

Accurately understanding human movement, specifically gait phase and the surrounding terrain, is crucial for developing adaptive lower-limb wearable robots like exoskeletons. These robots need to provide personalized assistance, which requires estimators to adapt quickly to new users and diverse walking conditions. However, existing methods struggle with:

  • Inter-subject and inter-terrain variability: People walk differently, and terrains vary greatly, making a one-size-fits-all approach ineffective.
  • Limited calibration data: In real-world scenarios, there's often only a small amount of data available for a new user or environment.
  • Nonlinearities of soft sensors: While comfortable and garment-integrable, fabric-based soft sensors introduce complexities like hysteresis and placement errors that complicate accurate estimation.

This paper aims to bridge this gap by developing a system that can learn to estimate gait and terrain features efficiently and accurately, even with limited data and the challenges posed by soft sensors.

Methods

The core of the proposed solution is a Model-Agnostic Meta-Learning (MAML) framework integrated with a Deep Convolutional Neural Network (DCNN).

  • DCNN Backbone: A multi-head DCNN is used to process time-series data from fabric sensors. It features a shared "feature extractor" that feeds into three separate "heads": one for locomotion mode classification, one for gait phase classification, and one for incline angle regression. This architecture allows for efficient parameter sharing and leverages the dependency between locomotion mode and joint kinematics by concatenating the mode prediction with features for the other heads.
  • MAML Integration: MAML is employed to learn a generalizable model initialization. This initialization is designed to capture subject- and terrain-invariant structures, enabling rapid adaptation to new tasks (i.e., new users or terrains) with only a few gradient steps.
  • Meta-Training: The DCNN is trained across multiple "tasks," where each task represents a specific subject and walking condition. During meta-training, the model undergoes task-specific adaptation on a "support set" and then updates its initial parameters based on the "query set" loss, aggregating gradients across all tasks. This process teaches the model how to learn efficiently.
  • Meta-Testing: For a new, unseen subject or terrain, the meta-trained model is fine-tuned on a small calibration "support set" from that specific task. This fine-tuning is significantly faster and requires less data than training from scratch.

Datasets

The study uses data collected from nine participants (7 male, 2 female) performing walking experiments under various conditions.

  • Terrains:
    • Treadmill: Level walking (LW), ramp ascent (RA), and ramp descent (RD) at different incline angles (0°, ±5°, ±10°) and speeds.
  • Staircase: Stair ascent (SA) and stair descent (SD) trials.
  • Sensors: Four fabric-based soft sensors embedded in compression pants at the hip and knee joints.
  • Ground Truth: Obtained from force plates on the treadmill and an inertial measurement unit (IMU) based finite-state machine (FSM) from ankle exoskeletons for stair trials. Locomotion modes and terrain geometries were labeled based on experimental protocols.

No new benchmark datasets were introduced; rather, existing data collection methods were adapted for evaluation.

Key Insights and Ideas

  • MAML for Soft Sensors: The paper demonstrates that MAML is highly effective in overcoming the nonlinearities and variability inherent in fabric-based soft sensors, a significant challenge for wearable sensing.
  • "Learning to Learn" for Adaptation: MAML's strength lies in learning an initialization that makes subsequent fine-tuning incredibly efficient. This is more powerful than simply pre-training on a large dataset, as it teaches the model how to adapt quickly to new data distributions.
  • Task-Specific Adaptation is Key: The meta-training process, which involves adapting to and aggregating gradients from diverse "tasks" (subject-terrain combinations), is crucial for achieving robust generalization.
  • Trade-offs in Baseline Performance: While standard pre-training methods (TL, SFT) can achieve good results, they often require more data and fine-tuning. Interestingly, training from scratch (RI) can sometimes outperform pre-trained models for locomotion mode classification when sufficient calibration data is available, highlighting the importance of data quantity for certain tasks.
  • Data Imbalance Mitigation: MAML shows resilience to data imbalance (e.g., more treadmill data than stair data), maintaining robust performance across different locomotion modes, unlike some baseline methods that exhibit performance degradation.

Key Results

The proposed MAML-based framework consistently outperformed baseline methods across various evaluation scenarios:

  • Data Efficiency (Scenario S1): MAML achieved the highest accuracy with the least amount of calibration data. As calibration data increased beyond 2 seconds, MAML consistently outperformed other approaches for all three estimation problems (gait phase, locomotion mode, incline angle).
  • Computational Efficiency (Scenario S2): With just one or two fine-tuning steps, MAML rapidly surpassed all baseline methods in accuracy. This highlights its superior adaptation speed.
  • Generalization (Scenario S3): In subject-specific and locomotion-mode-specific evaluations, MAML achieved the best average performance across all subjects and terrains.
  • Gait Phase Accuracy: MAML achieved an average accuracy of 0.85±0.01, significantly higher than DE (0.78±0.02), TL (0.79±0.02), and SFT (0.79±0.02).
  • Locomotion Mode Accuracy: MAML achieved a near-perfect average accuracy of 1.00±0.00 across all modes, while baselines varied.
  • Incline Angle RMSE: MAML achieved the lowest average RMSE of 2.67±0.09, outperforming baselines like DE (7.52±3.35) and SFT (2.40±0.65).

Limitations and Future Work

  • Limitations:
    • The study acknowledges that most gait phase misclassifications occur between adjacent phases, likely due to inherent ambiguity near transitions and hardware imperfections.
  • While MAML mitigates data imbalance, the performance differences between treadmill and stair scenarios for baselines suggest that further work might be needed to fully equalize performance across all terrain types.
  • Future Work:
    • The paper suggests exploring more complex terrain geometries and dynamic walking scenarios.
  • Investigating the integration of MAML with other sensor modalities or more advanced DCNN architectures could further enhance performance.
  • Expanding the evaluation to a larger and more diverse subject pool would strengthen the generalizability claims.

Summary updated

Abstract

This letter presents a model-agnostic meta-learning (MAML) based framework for simultaneous and accurate estimation of human gait phase and terrain geometry using a small set of fabric-based wearable soft sensors, with efficient adaptation to unseen subjects and strong generalization across different subjects and terrains. Inter-subject and inter-terrain variability, coupled with limited calibration data in real-world deployments, complicate accurate gait estimations. Further nonlinearities arise from fabric-based soft sensors, which improve comfort but introduce hysteresis, placement error, and fabric deformation. To address these challenges, the proposed framework integrates MAML into a deep learning architecture to learn a generalizable model initialization that captures subject- and terrain-invariant structure. This enables efficient adaptation to new users with a small amount of calibration data, while maintaining high estimation accuracy across subjects and terrains. Experiments on nine participants walking at various speeds over five terrain conditions demonstrate that the proposed framework outperforms baseline approaches in estimating gait phase, locomotion mode, and incline angle, with superior accuracy, adaptation efficiency, and generalization.
BibTeX
@article{zhu2026model,
  author = {Zenan Zhu and Wei Chen and Pei-Chun Kao and Janelle P. Clark and Lily Behnke and Rebecca Kramer-Bottiglio and Holly A. Yanco and Yan Gu},
  title = {{Model-agnostic Meta-learning for Adaptive Gait Phase and Terrain Geometry Estimation with Wearable Soft Sensors}},
  journal = {IEEE Robotics and Automation Letters},
  year = {2026},
  doi = {10.1109/lra.2026.3675944},
}
Figure