{"success":true,"data":{"author":"SAME-Limb authors (Harbin Institute of Technology)","bucketAccessType":"public","bucketName":"Tdongxu/A_Synchronized_Lower_Limb_AMG_sEMG_and_Mocap","bucketPrefix":"","category":"Physiology","cloudProvider":"huggingface","createdAt":"2026-11-22T01:09:18.629Z","description":"SAME-Limb, a dataset of time-aligned acceleromyography (AMG, muscle surface vibration measured by accelerometers), surface electromyography (sEMG), optical motion capture (MoCap) and knee and ankle joint angles recorded from 30 healthy adults (15 female, 15 male, mean age 23) at Harbin Institute of Technology. Subjects performed 11 functional lower-limb actions and walking at 1 to 5 km/h, four repetitions of 16 conditions, giving 1,918 released trials. Sixteen triaxial accelerometers (four at each of four left-leg muscle sites, 48 AMG axes) were interpolated to a common 1600 Hz grid, four matched EMG channels stay at 1000 Hz in microvolts, and 15 MoCap markers give right and left knee and ankle angle targets. Each trial is an NPZ file under data/trials/ with raw, DC-corrected and 5 Hz high-pass AMG and raw and preprocessed sEMG; CSV tables describe subjects, conditions, channels, auditory cues, excluded trials and cross-subject (20/3/7) and within-subject splits. The repository also ships the frozen benchmark code, 64 fitted joint-angle regression models and reference predictions. Uses: wearable muscle sensing, EMG versus AMG comparison, joint angle estimation for exoskeletons and rehabilitation. Data are CC BY 4.0 and code Apache-2.0, as the authors' LICENSING.md states. Cite the SAME-Limb paper, arXiv:2608.11958 (dataset DOI 10.57967/hf/9950).","displayTitle":"SAME-Limb: Synchronized Acceleromyography (AMG), Surface EMG and Motion Capture of Lower-Limb Muscles in 30 Adults (Harbin Institute of Technology)","doi":"10.57967/hf/9950","huggingFaceUrl":"https://huggingface.co/datasets/Tdongxu/A_Synchronized_Lower_Limb_AMG_sEMG_and_Mocap","id":"6473edf5-f391-54a1-a57e-0c1224d0ad40","license":"CC-BY-4.0","licenseEvidenceUrl":"https://huggingface.co/datasets/Tdongxu/A_Synchronized_Lower_Limb_AMG_sEMG_and_Mocap/raw/main/LICENSING.md","objectCountEstimate":2848,"probeStatus":"ok","referenceUrl":"https://arxiv.org/abs/2608.11958","region":"us","schemaCard":{"accessHints":[{"language":"python-datasets","snippet":"from datasets import load_dataset\nds = load_dataset(\"Tdongxu/A_Synchronized_Lower_Limb_AMG_sEMG_and_Mocap\")"},{"language":"python-hub","snippet":"from huggingface_hub import snapshot_download\nsnapshot_download(repo_id=\"Tdongxu/A_Synchronized_Lower_Limb_AMG_sEMG_and_Mocap\", repo_type=\"dataset\")"}],"caveats":["HuggingFace metadata: author=Tdongxu, license:cc-by-4.0, size_categories:1K<n<10K, format:csv, downloads=2152"],"format":"mixed","generatedAt":"2026-11-22T01:10:18.793Z","layout":{"example":"data/trials/user01/functional_action/deadlift/A1.npz","pattern":"{repoId}/{filename}"},"mountPath":"/mnt/archive","sampleHeaders":[{"bytes":"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","encoding":"hex","key":"data/trials/user01/functional_action/deadlift/A1.npz","truncatedAt":512}],"sampleObjects":["ATTRIBUTION.md","AUDITORY_CUE_METADATA.md","CITATION.cff","DATASET_CARD.md","DATA_AND_CODE_AVAILABILITY.md"],"sidecars":[{"bytes":"---\npretty_name: \"SAME-Limb: Synchronized AMG and EMG Dataset of Lower-limb Muscle Activities in Everyday Training\"\nlicense: cc-by-4.0\nlanguage:\n- en\ntags:\n- acceleromyography\n- surface-electromyography\n- motion-capture\n- biomechanics\n- time-series-regression\n- joint-angle-estimation\nsize_categories:\n- 1K<n<10K\nconfigs:\n- config_name: trials\n  default: true\n  data_files:\n  - split: train\n    path: metadata/trials.csv\n- config_name: subjects\n  data_files:\n  - split: train\n    path: metadata/subjects.csv\n- config_name: conditions\n  data_files:\n  - split: train\n    path: metadata/conditions.csv\n- config_name: channels\n  data_files:\n  - split: train\n    path: metadata/channels.csv\n- config_name: auditory_cues\n  data_files:\n  - split: train\n    path: metadata/auditory_cue_records.csv\n- config_name: excluded_trials\n  data_files:\n  - split: train\n    path: metadata/excluded_trials.csv\n- config_name: primary_benchmark\n  data_files:\n  - split: train\n    path: source_data/benchmark/primary_cross_subject_window_end.csv\n- config_name: benchmark_window_counts\n  data_files:\n  - split: train\n    path: source_data/benchmark/benchmark_window_counts.csv\n- config_name: model_inventory\n  data_files:\n  - split: train\n    path: models/model_inventory.csv\n---\n\n# SAME-Limb\n\n## Synchronized AMG and EMG Dataset of Lower-limb Muscle Activities in Everyday Training\n\n[![SAME-Limb overview: AMG and EMG sensing, joint-angle signals, and lower-limb activities](assets/same-limb-overview.png)](https://huggingface.co/datasets/Tdongxu/A_Synchronized_Lower_Limb_AMG_sEMG_and_Mocap/resolve/main/assets/same-limb-overview.png)\n\nThis publicly released dataset contains time-aligned acceleromyography (AMG),\nsurface electromyography (EMG), optical motion capture (MoCap), and four knee/ankle\njoint-angle signals from 30 subjects. The repository also provides the frozen\n5–100-Hz benchmark code, 64 fitted model artifacts, fixed reference\npredictions, source tables, and integrity manifests used for the associated\ndataset paper. SAME-Limb is the project and dataset name; the expanded article\ntitle above is retained verbatim for citation.\n\nProject website: [SAME-Limb](https://dongxutang918-afk.github.io/SAME-Limb/) - dataset overview, acquisition protocol, and benchmark introduction.\n\nRepository: <https://huggingface.co/datasets/Tdongxu/A_Synchronized_Lower_Limb_AMG_sEMG_and_Mocap>\n\nPersistent identifier: <https://doi.org/10.57967/hf/9950>\n\nAssociated article: <https://arxiv.org/abs/2608.11958>\n\nPaper DOI: <https://doi.org/10.48550/arXiv.2608.11958>\n\nRepository version: **2.0.3**\n\nVersion 2.0.3 corrects reader-facing metadata against arXiv:2608.11958v1 and\nadds the mixed-licence release boundary. The repository slug, trial NPZ keys,\nwindow-index filenames, window IDs, model weights, predictions, metrics, and\nfigure source values remain unchanged. See `RELEASE_NOTES_v2.0.3.md` and\n`metadata/naming_compatibility_v2.csv`.\n\n## Dataset Viewer configurations\n\nThe Dataset Viewer exposes nine intentionally separate tables: `trials`\n(default), `subjects`, `conditions`, `channels`, `auditory_cues`,\n`excluded_trials`, `primary_benchmark`, `benchmark_window_counts`, and\n`model_inventory`. Each configuration points to exactly one CSV file with its\nown schema. Other CSV, CSV.GZ, NPZ, model, prediction, metric, and source-data\nfiles remain downloadable repository assets but are not automatically merged\ninto the default Viewer split.\n\n## Dataset scope\n\n- 30 de-identified subjects (`user01`–`user30`);\n- 11 functional lower-limb actions and Walk at 1–5 km/h;\n- four planned repetitions for each of 16 conditions;\n- 1,918 released trials from 1,920 expected recordings;\n- 16 triaxial accelerometers, with four sensors at each of four muscle sites\n  (48 AMG axes);\n- four matched EMG channels;\n- 15 named MoCap markers;\n- AMG and EMG recorded only from the left leg, with bilateral MoCap-derived\n  knee and ankle labels;\n- four joint-angle outputs in the canonical order right knee, left knee,\n  right ankle, and left ankle;\n- cross-subject train/validation/test splits of 20/3/7 subjects;\n- with","encoding":"utf8","key":"README.md","name":"README.md","size":13124,"truncated":true},{"bytes":"# SAME-Limb introduction image\n\n- File: `same-limb-overview.png`\n- Source page: https://humitlab.github.io/download.html (SAME-Limb Database).\n- Original image: https://humitlab.github.io/assets/works/samelimbdatabase.png\n- Original size: 3264 x 889 pixels; PNG.\n- SHA-256: `3cda0868ae58b55f4485065bb8d2e9b6215b9ae55ff7f9afe5a88011201ec6e3`.\n- Reproduced byte-for-byte at the dataset owner's request; no resizing, cropping, recompression, or AI upscaling.\n\nThis image is credited to HUMIT Lab and included as a dataset introduction illustration. Copyright remains with the respective rightsholders. This patch grants no new licence to the image; the dataset's CC BY 4.0 and software's Apache-2.0 licences do not automatically apply to this credited website illustration. Consult the original source for permission to reuse it.\n","encoding":"utf8","key":"assets/README.md","name":"README.md","size":827,"truncated":false},{"bytes":"# Changelog\n\n## 2.0.3 repository patch\n\n- Use the paper display name `Forecasting (joint-angle label 100 ms after the\n  window end)` while retaining `ahead_100ms` and `future_100ms` as stable keys.\n- Use the paper frequency display names Full band, 20 Hz low-pass, 5–100 Hz\n  band-pass, and 100–760 Hz band-pass in release documentation.\n- Document the SAME-Limb name, associated arXiv article, physical units,\n  laterality boundary, synchronization statement, and mixed licensing.\n- Preserve all numerical algorithms, frozen keys, paths, weights, predictions,\n  metrics, and figure source values.\n\n## 2.0.1\n\n- Align reader-facing names with the submitted article: AMG, EMG, MoCap, and\n  target alignment.\n- Add `TARGET_ALIGNMENTS`; retain `TARGET_SETTINGS` as an exact compatibility\n  alias.\n- Record DOI `10.57967/hf/9950` and the five-author citation metadata.\n- Preserve all NPZ keys, index tokens, public paths, checkpoint bytes,\n  predictions, metrics, and figure source values from 2.0.0.\n\n## 2.0.0\n\n- Adopt the frozen complete-trial 5--100-Hz AMG representation.\n- Correct the AMG feature contract to channel-major statistics, including\n  mean absolute deviation from the channel mean and sum of squares.\n- Add native-rate Hudgins16 EMG features and AMG+EMG400 fusion.\n- Add the three frozen target-alignment and window-inventory contracts.\n- Compute MAE, RMSE, and Pearson r per output before four-output averaging.\n- Standardize the public API on subject, cross-subject, and within-subject\n  terminology.\n","encoding":"utf8","key":"code/CHANGELOG.md","name":"CHANGELOG.md","size":1519,"truncated":false},{"bytes":"# SAME-Limb AMG/EMG Joint-Angle Benchmark\n\nThis package accompanies the publicly released lower-limb AMG, surface\nelectromyography (EMG), MoCap, and joint-angle dataset. Version 2.0.1 preserves\nthe frozen\n5--100-Hz benchmark representation and the fixed reference-model inference\ncontract used for the submitted article while adding reader-facing naming\naliases. Repository version 2.0.3 updates the reader-facing Forecasting name\nto match arXiv:2608.11958v1. It does not change any numerical algorithm or\nfrozen artifact.\n\n## Primary benchmark representation\n\nThe public trial source key is `amg_dc_corrected_aligned` with shape\n`[samples, 48]` and nominal sampling rate 1600 Hz. The benchmark applies a\nzero-phase Butterworth 5--100-Hz band-pass filter to each complete trial before\nextracting any window. The SciPy design parameter is `N=4`; the band-pass\ntransformation yields four second-order sections. Filtering uses\n`scipy.signal.sosfiltfilt`, `axis=0`, odd extension, and `padlen=27`.\n\nThe package freezes the audited float64 SOS coefficients generated by\n`scipy.signal.butter(4, [5, 100], btype=\"bandpass\", fs=1600,\noutput=\"sos\")`. Freezing the coefficients prevents sub-ULP coefficient drift\nbetween compatible SciPy releases. The coefficient identity is checked before\nevery authoritative filtering call.\n\nAMG windows contain 480 samples from 48 channels (300 ms). On the acquisition\nclock, physical endpoints were generated with `np.arange(t[0] + 0.300,\nt[-1] + 1e-12, 0.100)` and mapped to sample bounds with\n`searchsorted(..., side=\"right\")`; the first frozen window is `[1:481)`.\nAbsolute acquisition timestamps are not distributed, so the released window\nindexes and their stored half-open sample bounds are the authoritative public\nidentities. They must not be regenerated from the trial-relative time arrays.\nThe three target alignments are:\n\n1. Window-end estimation: target at the physical window endpoint.\n2. Centered-window offline estimation: target 150 ms before the endpoint.\n3. Forecasting: joint-angle label 100 ms after the window end.\n\nThe released benchmark inventory contains 264,081 base window identities.\nAfter four-angle target validity checks, the respective totals are 263,999,\n263,994, and 263,833.\n\nUse `TARGET_ALIGNMENTS` for new code. `TARGET_SETTINGS` remains the same tuple\nas a backward-compatible alias. The published index tokens `endpoint`,\n`centered`, and `future_100ms` are intentionally unchanged.\n\n## Handcrafted features and modality fusion\n\n`extract_amg_features` returns 384 float32 features in channel-major order.\nFor each of 48 channels it computes mean, population standard deviation, root\nmean square, minimum, maximum, peak-to-peak range, mean absolute deviation\nfrom the channel mean, and sum of squares.\n\nThe EMG source key is `semg_preprocessed_basic`. The `semg_*` spelling is a\nstable published schema identifier. EMG remains at its native\n1000-Hz sampling rate. Each half-open 300-ms physical-time interval contains\n300 samples from four channels. `extract_emg_hudgins16` computes MAV, waveform\nlength, zero crossings, and slope-sign changes with epsilon zero and strict\ncomparisons, giving 16 features. `fuse_amg_emg_features` returns the frozen\n400-dimensional AMG+EMG representation.\n\n## Metric contract\n\n`regression_metrics` receives finite arrays with shape `[windows, 4]`. It\ncomputes MAE, RMSE, and Pearson r separately for right knee, left knee, right\nankle, and left ankle, then averages the four output-specific values. Valid\ntest windows are pooled before those four calculations. The function also\nreturns the four output-specific records so every aggregate can be audited.\n\n## Package modules\n\n- `filtering`: frozen 5--100-Hz design identity and complete-trial filtering;\n- `windowing`: acquisition-clock endpoint mapping, stored-index, native EMG\n  bound, target-frame, and inventory utilities;\n- `dataset`: safe public NPZ loading and canonical benchmark-index loading;\n- `features`: AMG384, EMG16, and AMG+EMG400 extraction;\n- `metrics`: output-wise MAE, RMSE, and Pearson r aggregation;\n- `models`: TCN, LSTMPlus","encoding":"utf8","key":"code/README.md","name":"README.md","size":4604,"truncated":true}],"sourceObjectKey":"data/trials/user01/functional_action/deadlift/A1.npz","starterCell":"from huggingface_hub import hf_hub_download\npath = hf_hub_download(repo_id=\"Tdongxu/A_Synchronized_Lower_Limb_AMG_sEMG_and_Mocap\", filename=\"data/trials/user01/functional_action/deadlift/A1.npz\", repo_type=\"dataset\")\nprint(path)","starterSamplePath":"ATTRIBUTION.md","totalObjects":2848},"summary":"SAME-Limb, a dataset of time-aligned acceleromyography (AMG, muscle surface vibration measured by accelerometers), surface electromyography (sEMG), optical motion capture (MoCap)…","title":"SAME-Limb: Synchronized Acceleromyography (AMG), Surface EMG and Motion Capture of Lower-Limb Muscles in 30 Adults (Harbin Institute of Technology)","updatedAt":"2026-11-22T04:04:26.965Z","visibility":"public","websiteUrl":"https://dongxutang918-afk.github.io/SAME-Limb/"}}