SAME—Limb.
HUMIT LabHarbin Institute of Technology, Shenzhen

SAME-LIMB · OPEN RESEARCH DATASET

Synchronized AMG and EMG Dataset of Lower-limb Muscle Activities in Everyday Training

Dongxu Tang ↗ · Shih Ying-Lei · Zhuoyi Ren · Jianting Liao · Yitian Shao

HUMIT Lab · Harbin Institute of Technology, Shenzhen

Synchronized acceleromyography, surface electromyography, and optical motion capture from 30 participants across 16 lower-limb task conditions.

Read the paper Access dataset
AMG · 1,600 HzEMG · 1,000 HzMoCap · 90 Hz
1,918 trials · 4 muscle sites · bilateral knee and ankle references
SENSING & OPTICAL MOTION CAPTURE
Loading the sensing system
NOKOV Mars2H · 16 cameras · 90 Hz
Hover over the model to explore the sensors.

Drag horizontally to rotate · Arrow keys when focused

Illustrative walking · Arms crossed over the chest
Physical-scale layout · Selected sensor enlarged for inspection

Dataset overview

THE DATASET

Dataset overview

SAME-Limb brings together acceleromyography, surface electromyography, and optical motion capture from everyday lower-limb training—a reproducible foundation for studying muscle activity and estimating joint angles.

30healthy participants15 female · 15 male
1,918released trialsof 1,920 planned recordings
16task conditions11 functional tasks · 5 walking speeds
4joint-angle outputsBilateral knees and ankles
01 / MECHANICAL

Acceleromyography
16 triaxial sensors

Sixteen AIS2IHTR triaxial accelerometers capture local mechanical signals at four sites on the left leg.

1,600 HzNominal processing grid
02 / ELECTRICAL

Surface EMG
4 matched channels

Four LE230 surface EMG channels measure electrical activity at the matched RF, VM, TA, and GA muscle sites.

1,000 HzNative sampling · µV
03 / KINEMATICS

Optical motion capture
NOKOV Mars2H

16 infrared cameras track 15 markers to provide reference angles for both knees and ankles.

Left-leg sensing. Bilateral references. AMG and EMG were recorded only from the left leg. Right-leg angle estimates are not direct measurements of right-leg muscle activity. Joint angles are geometric labels defined by markers and segment vectors.

ACQUISITION PROTOCOL

Standardized lower-limb activities

Eleven functional actions and five walking speeds. Select an activity for its essential instructions.

Supplementary Figure S1: all 11 functional lower-limb actions and treadmill walking, shown together in the original study photographs
Supplementary Fig. S1 · All recorded activities Open full figure

Deep squat

Use the same foot placement and height-adjusted box setup as the deadlift, with no standardized load inside. Complete the squat to the auditory cues.

Auditory cues at approximately 2, 6, and 10 s.
Arms crossed during functional actions, unless hand contact is required.Full protocol · SI Table S2 ↗

REPRODUCIBLE BY DESIGN

Benchmark and ablation results

Compare model performance, task-specific errors, and acquisition configurations using the published source data.

BENCHMARKSABLATION EXPERIMENTS

Table V · Pooled test-window metrics

Cross-subject model performance

8.840°TCN · lowest primary MAE
20 / 3 / 7Training / validation / test subjects
300 msInput window · 100 ms stride
MAE (°) ↓ lower is better

Seed 42; one training run per configuration. Metrics are computed over valid pooled test windows for each of the four joint outputs, then averaged across outputs.

Numerical values
ConfigurationMAE (°)RMSE (°)Pearson r
Random Forest9.59113.6390.634
TCN8.84012.5150.672
LSTMPlus9.53513.5890.630
InceptionTimePlus9.42612.9320.658
Paper Table V · Pooled test-window metrics
CSVSource

FROM SIGNAL TO MOVEMENT

Representative joint-angle predictions

Representative deep-squat, vertical-jump, and 1 km/h walking trials. Black traces show MoCap labels, dashed blue traces Random Forest, and pink traces TCN.

Figure 4: left-knee, left-ankle, right-knee, and right-ankle references and predictions for three representative trialsFig. 4 · Open original

Representative trials complement the aggregate evaluation; they do not replace it. Source: arXiv:2608.11958v1, Fig. 4.

READY FOR YOUR RESEARCH

Data access and reproducibility

Get the dataset

Complete trial data

1,918 NPZ files with AMG, EMG, MoCap, joint angles, validity masks, and modality-specific time vectors.

Browse trials

Reproducible benchmarks

Fixed splits, window manifests, 64 fitted model artifacts, reference predictions, and source tables.

Explore benchmark code

Documented timing

AMG streams share a common grid. All modalities are cropped to their shared interval; labels use the nearest native MoCap frame.

Read the data guide
Data format and synchronization

The four AMG streams are linearly interpolated onto a nominal 1,600 Hz grid, without extrapolation. EMG retains its native 1,000 Hz sampling rate; MoCap retains native frame timestamps. Device clocks were coordinated by NTP, with no additional correction for cross-device clock offset or drift. Hardware-level synchronization accuracy was not established.

NPZ arrays include amg_raw_aligned, amg_dc_corrected_aligned, amg_highpass_5hz_aligned, semg_raw, semg_preprocessed_basic, motion_markers_named, motion_skeleton_selected, joint_angles_deg and joint_angle_valid_mask, provided alongside modality-specific time vectors and metadata_json. Raw AMG values are stored in g/digit; multiply by 19.14258 to convert to mm/s².

Auditory cue records are available for 1,726 trials and absent for 192. They were not used for alignment or benchmark computation.

Scope, ethics, and licensing

Participants were healthy adults aged 23.4 ± 2.1 years. Clinical generalization to patients or other populations has not been validated. The study was approved by the Medical Ethics Committee of Harbin Institute of Technology (HIT-2024046; 8 July 2024). All participants provided written informed consent.

Data, metadata, research source materials, predictions, metrics, and trained models are licensed under CC BY 4.0; code/ software uses Apache-2.0. Third-party materials retain their original terms. The article and Supplementary Information have separate publication licenses.

Read the full licensing terms ↗
Interactive anatomy and hardware references

The model uses a CC0 MakeHuman-derived mesh with an illustrative walking cycle and arms crossed over the chest. Transparent left-leg skin reveals schematic RF, VM, TA, and GA muscles. Fifteen reflective markers follow the lower-body segments; positions are schematic rather than measured participant coordinates.

LE230 geometry follows Biometrics dimensions: 42 × 24 × 14 mm. The AIS2IHTR IC measures 2 × 2 × 0.93 mm; its custom carrier board is illustrated at approximately 6 × 5 mm, estimated from paper Fig. 1. Devices use these dimensions in the physical-scale layout; selection temporarily enlarges one sensor for inspection.

Product reference images belong to Biometrics Ltd and STMicroelectronics. The human mesh was distributed through MakeHuman / MPFB2, NAVER anny, and three.ws.

View model license and hardware sources ↗

THE PAPER

arXiv:2608.11958v1 · 12 AUG 2026

Synchronized AMG and EMG Dataset of Lower-limb Muscle Activities in Everyday Training

Dongxu Tang · Shih Ying-Lei · Zhuoyi Ren · Jianting Liao · Yitian Shao

School of Computer Science and Technology
Harbin Institute of Technology, Shenzhen

Read the paper Download PDF
Cite this work
@misc{tang2026synchronizedamgemgdataset,
      title={Synchronized AMG and EMG Dataset of Lower-limb Muscle Activities in Everyday Training}, 
      author={Dongxu Tang and Shih Ying-Lei and Zhuoyi Ren and Jianting Liao and Yitian Shao},
      year={2026},
      eprint={2608.11958},
      archivePrefix={arXiv},
      primaryClass={cs.HC},
      url={https://arxiv.org/abs/2608.11958}, 
}