# MoCap-to-IMU Transfer Learning for Industrial Exoskeletons
> Explore a teacher-student transfer learning pipeline for industrial exoskeleton intent prediction using knowledge distillation from MoCap to IMU sensors.

Tags: exoskeleton, transfer-learning, knowledge-distillation, imu-sensors, mocap, robotics, industrial-automation, ai
## Robust MoCap-to-IMU Transfer for Industrial Exoskeleton Intent Modeling Under Domain Shift
- **Project:** EMPOWER Project — Industrial Human-Robot Collaboration.
- **Presenter:** Alba Huti (MSc Thesis Proposal).

## Problem Statement
- **Real-time Intent Prediction:** Current models rely on lab-grade sensors that are not scalable.
- **Gaps identified:** Lack of standardized pipelines for transferability, performance degradation under domain shift (robustness), and low industrial acceptance due to usability concerns.

## Research Objective
- Development of a **Teacher-Student transfer learning pipeline**.
- **Teacher Model:** High-capacity MS-G3D skeleton graph network trained on lab-only MoCap kinematics.
- **Student Model:** Compact TCN/LSTM architecture using 2–4 wearable IMUs for industrial edge deployment.

## Methodology
- **Datasets:** LARa Dataset (MoCap + IMU) and OpenPack Dataset (IMU only).
- **Framework:** Knowledge distillation using distillation loss (KL) and feature alignment (MSE).
- **Target:** Edge-feasible real-time inference support using only acceleration and gyroscope data.

## Analysis & Evaluation
- **Stress Tests:** Cross-subject generalization, cross-task/scenario shift, sensor rotation/noise, and cross-dataset testing.
- **Key Metrics:** Macro-F1, Accuracy, RMSE/MAE, Latency (ms), and Model Size.

## Expected Contributions
- Reproducible MoCap-to-IMU transfer pipeline.
- Guidance on minimal sensor configurations (identifying the trade-off curve between sensor count and performance).
- Empirical evidence of improved robustness under domain shift via privileged supervision.
- Standardized benchmarking protocol for the field.
---
This presentation was created with [Bobr AI](https://bobr.ai) — an AI presentation generator.