MoCap-to-IMU Transfer Learning for Industrial Exoskeletons

Robust MoCap-to-IMU Transfer for Industrial Exoskeleton Intent Modeling Under Domain Shift

MSC THESIS PROPOSAL

Alba Huti

EMPOWER Project — Industrial Human-Robot Collaboration

Problem Statement

Industrial exoskeletons need real-time intent prediction — but current approaches rely on lab-grade sensors that don't scale to real-world deployment.

Transferability Gap

Current research relies on MoCap & EMG; no standardized pipeline exists for transferring knowledge to IMU-only wearable deployment models.

Robustness Gap

Performance degrades under domain shift (across users, tasks, sensor configs) but rarely addressed with explicit mitigation strategies.

Acceptance Gap

Many studies under-measure usability; industrial adoption depends on comfort, perceived usefulness, and predictable behavior.

EMPOWER Project — Industrial Human-Robot Collaboration

Research Objective

To develop a robust teacher–student transfer learning pipeline that distills MoCap-based kinematic knowledge into a compact, IMU-only model for real-time exoskeleton intent prediction in industrial settings.

Teacher Model

High-capacity temporal network trained on full-body MoCap kinematics (MS-G3D skeleton graph).

Rich Supervision — Lab Only

Student Model

Compact TCN/LSTM using 2–4 wearable IMUs only.

Edge Deployment — Industrial Ready

Knowledge Distillation

Cross-Modality Transfer

Edge Feasibility

EMPOWER Project — Industrial Human-Robot Collaboration

Research Questions

RQ1

Can MoCap data, used as privileged supervision during training, improve an IMU-only model's performance compared to a model trained solely on wearable IMU data?

RQ2

Can a model using only minimal wearable IMU sensors (2–4) accurately predict user intent or generate outputs relevant for exoskeleton support and assistance?

RQ3

Can laboratory-quality kinematic information from MoCap be effectively distilled into a practical IMU-based system while maintaining robustness under real-world domain shifts (unseen users, scenarios, sensor placement variation, noise)?

Methodology

EMPOWER Project — Industrial Human-Robot Collaboration

Proposed Analysis Tools & Evaluation

EMPOWER Project — Industrial Human-Robot Collaboration

Calendar of Activities — Gantt Chart

Activity

M1

M2

M3

M4

M5

M6

M7

M8

M9

1. Literature Review & Dataset Acquisition

2. Data Preprocessing & Baseline Implementation

(IMU-only CNN+LSTM)

3. Teacher Model Training

(MS-G3D on MoCap)

4. Knowledge Distillation Framework

(Student training)

5. Robustness Evaluation

(cross-user, cross-task, sensor perturbations)

6. Deployment Feasibility & Edge Testing

7. Writing & Thesis Submission

Expected Contributions

Reproducible MoCap-to-IMU Transfer Pipeline

A concrete teacher–student distillation framework for temporal motion data, directly operationalizing key literature recommendations for industrial exoskeleton intent modeling.

Minimal Sensor Configuration Guidance

A clear performance vs. sensor-count trade-off curve identifying the "minimal sufficient" IMU setup for robust industrial deployment.

Robustness Under Domain Shift

Empirical evidence that MoCap-based privileged supervision improves not only peak accuracy but also generalization stability across users, tasks, and sensor perturbations.

Standardized Benchmarking Protocol

Clearly defined cross-user/cross-scenario splits, synthetic perturbations, and metrics (Macro-F1, RMSE, Latency, Model Size) to address evaluation fragmentation in the field.

Deployment-Oriented Insights

Quantitative links between sensor count, robustness, latency, and model size — transforming research into practical design guidance for embedded exoskeleton systems.