SoftRigid Coupling Exoskeleton

刚柔并"脊"——新一代万向刚柔耦合仿生外骨骼运动预判系统

> College Students' Innovation and Entrepreneurship Competition Project | Real-time human motion prediction for soft-rigid coupled exoskeletons
> https://github.com/ShBWei/SoftRigid-Coupling-Exoskeleton/

Overview

A physics-guided deep learning system for real-time prediction of spinal posture and rigidity states in segmented exoskeletons.

• Input: 12 IMU streams @ 60Hz, 24-frame (400ms) observation window
• Output: 12-frame (200ms) future prediction of posture + mode + switch probability
• Response target: delay under 200ms (aligned with plan section 3.2.4)

12-IMU Configuration

We use a 12-node full-body kinematic chain instead of the sparse 8-node setup to enable closed-loop drift correction and bilateral symmetry constraints.

Node	Location	Role	Channels
0	Cervical (C4-C6)	Head-neck reference	6
1	Thoracic (T6-T8)	Thoracic kyphosis monitoring	6
2	Lumbar (L2-L4)	Lumbar lordosis core	6
3	Sacral (S1-S3)	Fixed base, zero-drift anchor	6
4	Left Shoulder	Upper limb chain start	6
5	Right Shoulder	Bilateral pair	6
6	Left Elbow	Elbow flexion axis	6
7	Right Elbow	Bilateral pair	6
8	Left Hip	Lower limb chain start	6
9	Right Hip	Bilateral pair	6
10	Left Knee	Knee flexion axis	6
11	Right Knee	Bilateral pair	6
Total			72 ch/frame

Input tensor: 12 x 6 x 24 = 1728 dimensions.

Four-Segment Spine Structure

[Cervical x7] -- [Thoracic x12] -- [Lumbar x5] -- [Sacral fused]
   C1-C7             T1-T12            L1-L5         fixed base

The sacral node serves as the kinematic root to eliminate integration drift. Bilateral shoulder/elbow/hip/knee pairs enforce symmetry consistency during training.

Model Outputs

• Posture head: 23 x 3 joint angles
• Mode head: rigid/flexible classification (0/1)
• Switch head: switching probability (0-1)
• Physics layer: differentiable soft ROM constraints + hard clipping + bilateral symmetry loss

Data and Validation

Current release uses synthetically generated biomechanical data for algorithm closed-loop verification. The data is constrained by anatomical ROM (ISB coordinate system), four-segment spine coupling covariance, bilateral symmetry (< 3 deg deviation), and inverted-pendulum gait templates.

Scenario distribution in the validation set:

Scenario	Samples	Ratio	Mode
Normal gait	360	60%	flexible
Scoliosis simulation	60	10%	flexible -> rigid
Load lifting	120	20%	rigid
Fall-prevention	60	10%	transition

Reference baseline: The reported cohort effect (ROM +42%, burden -65%) is derived from published meta-analyses on exoskeleton-assisted rehabilitation, not from an independent clinical trial of this project.

Training Metrics

• MPJPE (pose error proxy)
• Mode-switch accuracy / F1
• Ratio of response delay < 200ms
• ROM violation rate
• Bilateral symmetry error

Quick Start

git clone https://github.com/ShBWei/SoftRigid-Coupling-Exoskeleton.git
cd SoftRigid-Coupling-Exoskeleton
pip install -r requirements.txt

# Edit config.py to set IMU nodes, sample size, epochs, etc.
# Then run the full pipeline:
python main.py

Generated outputs in results/:

• medical_data_12imu.npz — synthetic dataset
• baseline_best.pth, ours_best.pth — model weights
• learning_curves.png
• fig1_spine_segment.png
• fig2_symmetry.png
• fig3_fall_timeline.png
• table_comparison.tex

Comparison Snapshot (Lightweight Run)

Experiment	Metric	Baseline-8	Baseline-12	Ours-12
Normal gait	MPJPE (deg)	3.24	2.74	2.22
Scoliosis	MPJPE (deg)	3.85	2.89	2.45
Load lifting	ROM violation (%)	1.87	0.43	0.00
Fall-prevention	Response delay (ms)	198	185	167

Note: This is a quick end-to-end verification run. Paper-grade significance requires repeat_seeds>=5 and longer training (epochs>=30).

Project Structure

SoftRigid-Coupling-Exos/
├── results/                    # Output directory (figures, tables, checkpoints)
├── config.py                   # Hyperparameters and IMU configuration
├── data_generator.py           # Biomechanical synthetic data generation
├── main.py                     # End-to-end pipeline entry
├── training_framework.py       # Model training, validation, and testing
├── visualization.py            # Result plotting and figure export
├── requirements.txt
├── LICENSE
└── README.md

Disclaimer

1. Project nature: This is a college student innovation and entrepreneurship training program project for algorithm exploration. It does not constitute a medical device, clinical diagnostic tool, or therapy system.
2. Data: All current results are based on synthetic biomechanical data for architecture validation. No clinical cohort has been enrolled or tested.
3. Performance: The metrics above are from lightweight training configurations and lack statistical significance for clinical claims.
4. Safety: Any hardware deployment of exoskeleton control requires mechanical safety assessment, ethical review (IEC/IRB), and real-time redundancy testing. The authors assume no liability for direct or indirect use.
5. Citation: If referenced in academic or competition contexts, please state clearly that the algorithm is in research preview and results are simulation-based.

Citation

@misc{softrigid2026,
  title={SoftRigid Coupling Exoskeleton: A 12-IMU Full-Body Kinematic Chain with Physics-Guided Spatiotemporal Networks},
  author={ShBWei},
  year={2026},
  howpublished={\url{https://github.com/ShBWei/SoftRigid-Coupling-Exoskeleton/}},
  note={College Student Innovation and Entrepreneurship Training Program. Results based on synthetic biomechanical data.}
}

License

This project is licensed under the MIT License.