RehabMotion-KIMORE-Finetuned

Synthetic rehabilitation movement dataset generated with MotionGPT3 fine-tuned on the KIMORE dataset. Part of ongoing research at UFES on augmenting rehabilitation datasets with transformer-based motion generation.

This is the fine-tuned release — see the companion RehabMotion-KIMORE-Baseline dataset for the pre-trained baseline.

Dataset Summary

• Samples: 1000 motion sequences
• Exercises: 5 rehabilitation exercises (KIMORE protocol)
• Format: SMPL 22-joint skeleton, 20 fps, positions in meters
• Generator: MotionGPT3 fine-tuned on the KIMORE rehabilitation dataset (200 epochs, AdamW, lr 5e-6, RTX 4090)
• Conditioning: text prompts describing each rehab exercise

Exercise Distribution

Exercise	Samples
sagittal_arm_lift	200
lateral_trunk_tilt	200
trunk_rotation	200
pelvis_rotation	200
squatting	200

Data Structure

RehabMotion-KIMORE-Finetuned/
├── README.md                # This dataset card
├── LICENSE                  # CC-BY-NC-4.0
├── metadata.json            # Per-sample: filename, exercise, prompt, n_frames, shape
├── evaluation_results.json  # Biomechanical metrics vs real KIMORE
├── prompts.txt              # All text prompts used for generation
├── motions/                 # .npy files, shape (1, n_frames, 22, 3) — SMPL joint positions
└── features/                # .npy files, shape (n_frames, 263) — HumanML3D features

Loading a sample

import numpy as np
from huggingface_hub import hf_hub_download

# Download one motion
path = hf_hub_download(
    repo_id="lucasbrandao/RehabMotion-KIMORE-Finetuned",
    filename="motions/synthetic_squatting_00400.npy",
    repo_type="dataset",
)
motion = np.load(path)  # shape: (1, n_frames, 22, 3)
print(motion.shape)

Joint order (SMPL-22)

pelvis, l_hip, r_hip, spine1, l_knee, r_knee, spine2, l_ankle, r_ankle, spine3, l_foot, r_foot, neck, l_collar, r_collar, head, l_shoulder, r_shoulder, l_elbow, r_elbow, l_wrist, r_wrist

Exercises (KIMORE Protocol)

ID	Name	Target
Es1	Lateral arm elevation	Shoulder
Es2	Arm flexion with elbows at hips	Elbow
Es3	Trunk rotation (seated)	Thoracic spine
Es4	Pelvis rotation (standing)	Lumbar spine
Es5	Squatting	Lower limbs

Quality Metrics

Biomechanical evaluation over all 1000 synthetic samples (and distribution comparison against the real KIMORE dataset when available):

Metric	Value
Jerk (mean, lower = smoother)	20.20 ± 8.41
Duration (s)	5.53 ± 0.92
Biomechanical validity (frames within elbow/knee ROM, min across joints)	94.8%
Synthetic mean joint velocity (m/s)	0.2627
Real KIMORE mean joint velocity (m/s)	0.2397
KL divergence (real ‖ synthetic velocity)	0.0471

Synthetic motions are smoother than Kinect-captured ones because the generator produces clean trajectories (no sensor noise). Biomechanical validity is the share of frames whose elbow and knee angles fall inside textbook range of motion (elbow interior angle 30-180 deg, knee 40-180 deg); per-joint figures are in evaluation_results.json.

Generation Details

• Model: MotionGPT3 fine-tuned on KIMORE (MoT architecture, GPT2 backbone + VAE + diffusion head)
• Fine-tuning: 200 epochs on KIMORE (~1963 train samples), AdamW lr 5e-6, cosine schedule, batch 16
• Hardware: NVIDIA RTX 4090 24GB (~3.5h total fine-tuning)
• Best checkpoint selected by validation FID
• Task: Text-to-Motion (t2m)
• Inference: GPU, ~3s per sample
• Prompts: 10 text-template variations per exercise

See source repo for generation scripts and full pipeline.

Intended Use

• Research only (CC-BY-NC-4.0): academic study of synthetic motion augmentation for rehabilitation analysis, movement classification, and generative model benchmarking.
• Not for clinical use: these motions are generated by a generic motion model and have not been validated by physiotherapists. Do not use them for diagnosis, therapy planning, or patient-facing applications.

Limitations

• Healthy-like motions: although fine-tuned on KIMORE (which contains both control and pathological groups), the model still tends to produce reasonably healthy motions and may not capture all patient-specific compensations or restricted ROM.
• Smoother than reality: synthetic jerk is typically lower than sensor-captured motion because the generator produces clean trajectories (no Kinect noise).
• Limited prompt diversity: 50 base prompts (10 per exercise).

Citation

If you use this dataset, please cite:

@misc{physiomotion-synthetic-finetuned,
  title  = {RehabMotion-KIMORE-Finetuned: Rehabilitation Motion Dataset via MotionGPT3},
  author = {Brandão, Lucas Dantas},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/lucasbrandao/RehabMotion-KIMORE-Finetuned}},
  note   = {UFES research project}
}

And the upstream works:

@article{motiongpt3,
  title   = {MotionGPT3: Human Motion as a Second Modality},
  journal = {arXiv:2506.24086},
  year    = {2025}
}

@article{kimore,
  title   = {KIMORE: Kinematic Assessment of Movement and Clinical Scores
             for Remote Monitoring of Physical Rehabilitation},
  author  = {Capecci, M. et al.},
  journal = {IEEE Transactions on Neural Systems and Rehabilitation Engineering},
  year    = {2019}
}

License

Released under CC-BY-NC-4.0. You may share and adapt the material with attribution, for non-commercial purposes only.