Tagline. Generate hip & knee joint trajectories from user/biomech features using a Deep Neural Network (key-points → spline) and Gaussian Process Regression (gait period/phase).
Snapshot. Mean relative errors on full trajectories: knee 5.3%, hip flex. 5.8%, hip abd. 9.8%; gait-period RMSE ≈ 0.054 s.
Method at a glance — key-points → spline trajectory → absolute time via GPR.
Robotic exoskeletons need smooth, reliable joint trajectories that adapt to the user and context (e.g., speed, incline).
This repository shares a clean project structure, data instructions, and results of a high-level ML controller:
- DNN predicts joint trajectory key-points (hip/knee) → cubic spline interpolation.
- GPR estimates gait period/phase to map trajectories into absolute time.
No private data is included. See
data/README.mdfor the datasets, how to obtain them legally, and how to prepare your own files or use the synthetic samples.
exoskeleton-high-level-control-ml/
├─ data/
│ ├─ README.md
│ └─ samples/
├─ docs/
│ └─ figures/
├─ research/
├─ simulation_models/
├─ src/
│ └─ exo_ml/
├─ .gitignore
└─ README.md
-
Clone bash git clone https://github.com/AidenPQ/exoskeleton-high-level-control-ml.git cd exoskeleton-high-level-control-ml
-
(Optional) Environment python -m venv .venv
source .venv/bin/activate pip install -U pip
-
(Optional) Minimal deps — for quick plots/metrics: pip install numpy pandas matplotlib scikit-learn scipy
-
Data Read data/README.md: official DOIs/URLs, HDF5 schema (datasetV2.h5), and the CSV mapping for training/eval.
Use the synthetic samples in data/samples/ for quick illustration if you don’t have the real datasets yet.
Data (what matters)
No raw data is redistributed here.
data/README.md documents three public datasets (Embry 2018; Fukuchi 2018; Moreira 2021), the subject → trial → side → angles HDF5 structure (+ speed/incline metadata & demographics), and a flat CSV export recommended for models.
Angle matrices are normalized per gait cycle (150 points/cycle).
Minimal CSV columns recommended: subject_id, trial_id, side, joint, axis, cycle_idx, cycle_pct, value_deg, speed_mps, incline_deg, age, gender, height_m, weight_kg.
Key-point → Spline (degree 3)
Relative RMSE (mean ± std): knee 1.8% ± 1.2%, hip flex. 2.1% ± 3.9%, hip abd. 3.7% ± 2.6%
Vast majority < 6% (rare outliers).
Full trajectory generation (modified complete model)
Mean relative RMSE: knee 5.3%, hip flex. 5.8%, hip abd. 9.8%
Max rel. RMSE observed (transparency): knee 0.21, hip flex. 0.72, hip abd. 1.04
Outliers > 10% markedly reduced vs initial model (~< 7%).
Gait period (GPR, Matérn 1.5)
RMSE test ≈ 0.054 s (≈ 1.35%–10% relative depending on the period).
GPR — predicted gait period vs speed (95% band).
Interpolation error from key-points for Hip flexion(relative RMSE distribution).
Final trajectory interpolation for Hip flexion.
(Optional transparency) You can also include worst/best case plots in docs/figures/worst_cases_modified.png and docs/figures/best_cases_modified.png.
Key-points: DNN predicts a compact set of joint key-points (hip/knee).
Spline: a cubic spline reconstructs a continuous trajectory over the normalized gait cycle (0–100%).
GPR: a Gaussian Process (Matérn 1.5) predicts the gait period/phase, mapping the trajectory to absolute time.
Evaluation: relative RMSE per joint + best/worst cases for transparency.
No license is granted. All rights reserved by the author. If you wish to reuse parts of the code, figures, or methodology, please open an Issue to request permission.
If this repository informs your work, please cite as:
Ngandjui Tiako, P. Q. (2024–2025). High-Level Control for Lower-Limb Exoskeletons — DNN + GPR. Repository: AidenPQ/exoskeleton-high-level-control-ml.
LinkedIn — Pierre Quentin NGANDJUI TIAKO: https://www.linkedin.com/in/pierre-quentin-ngandjui-tiako-006975180/