Research and data

Open Data,
Baselines Against a Constant

Forearm force-myography recordings with continuous finger flexion and extension labels (per-frame regression targets, not gesture classes), built for researchers in human-machine interfaces, prosthetic sensing and VR interfaces. OpenMuscle also publishes its analysis code and a technical report (v0.3) that measures how well finger prediction from the forearm generalizes across sessions and wearers, each result next to a constant baseline. Everything is free to use under open licenses.

Where this stands: the report is a preprint draft (v0.3, 2026-11-09) and has not been peer reviewed. Its main analysis covers two wearers and four sessions on one hardware revision (FlexGrid V4), with a smaller legacy subset from a single V3 band, so with two wearers these are observations, not general findings. The numbers below are reported as they are, including the unflattering ones.

The dataset

Labeled and raw sensor capture sessions from OpenMuscle wearables, published for machine learning research on hand and forearm biomechanics. The data is free to use, including commercially, under CC BY 4.0 with attribution to the OpenMuscle project. It grows as new sessions are recorded.

What is in the dataset

Config Captures Board Labels Recorded
lask5_labeled default 23 FlexGrid V3 (20), earlier pre-V4 board not recorded (3) LASK5 handheld labeler, 4 piston targets May 2026
quest_labeled 25 FlexGrid V4 (14), FlexGrid V3 (11) Meta Quest hand tracking, 25 joints per hand May to July 2026
legacy_2025 13 2025 prototype band LASK5, 4 piston targets, partial coverage July 2025
supplementary_unlabeled 14 2025 prototype band (13), earlier pre-V4 board not recorded (1) None (long unlabeled streams) July 2025 to May 2026

75 captures in all.

Wearers: the dataset codes the two wearers as P01 and P02. The paper calls them W1 and W2.
Splits: evaluate with a temporal split or leave-one-session-out. Random row splits overstate skill, because neighbouring frames are near-duplicates (see the baseline results).

What the GitHub archive holds

Folder Contents
flexgrid/ FlexGrid prototype capture sessions from 2025 (CSV, 60 or 64 pressure cells; the GitHub README labels them V3)
merged/ Merged multi-session training CSVs produced by the PC pipeline
curated/ Hand-picked good sessions from early capture runs
legacy/ Early prototype captures (txt) from pre-FlexGrid devices

The repository README is the authoritative description of what is published today.

Device and signal (as used in the paper)

Band FlexGrid V4, ESP32-S3 controller on a flexible sensor sheet
Sensing 15 x 4 grid, 60 Velostat pressure cells, about 12-bit readings
Rate and link 17 to 19 Hz per band, streamed over Wi-Fi UDP
Motion 6-axis IMU on the band
Labels Meta Quest hand tracking (25-joint poses) on V4. The LASK5 handheld piston labeler was used on the earlier V3 band; no V4 recording is LASK5-labeled yet

The paper's analysis set

Wearers 2
Sessions 4 (8 two-hand captures)
Paired sensor-label frames about 22,300 (11,145 left hand, 11,146 right hand)
Targets 15 wrist-relative finger degrees of freedom per hand

Licenses

Dataset CC BY 4.0
Hardware CERN-OHL-S-2.0
Software MIT

The paper

Preprint draft, technical report v0.3,

OpenMuscle: An Open Forearm Force-Myography Dataset with Continuous Finger Labels, and Baselines Across Sessions and Wearers

Tory Moghadam, OpenMuscle, 2026

The report presents an open system that pairs low-cost 60-cell forearm force-myography bands with two sources of continuous finger labels, consumer VR hand tracking and the LASK5 handheld labeler. It releases the recordings and reports frame-wise baselines under a four-rung generalization ladder, from a random split within a session up to prediction across wearers, each next to a constant baseline that predicts the training-mean pose.

Every number in the main text regenerates from the public release with analysis/paper/analyze.py and figures.py (see Reproduce it).

Baseline results: the generalization ladder

One deliberately simple model, evaluated four ways, each time next to a constant baseline that predicts the training-set mean pose for every test frame. A model that scores worse than that constant has learned nothing that transfers. Error is the mean finger-flexion MAE as a fraction of full curl, averaged over runs. Lower is better.

Rung Protocol Runs Model flexion MAE Constant baseline Model flexion R²
E1 Within one session, random 80/20 split 8 0.043 4.9 deg per joint 0.197 0.84 to 0.91
E2 Within one session, temporal split (train on the first 80% of the timeline) 8 0.110 12.7 deg per joint 0.218 -2.37 to 0.77, negative in 2
E3 Across sessions, same wearer (train on one session, test on another) 8 0.244 25.7 deg per joint worse than constant 0.203 -2.34 to 0.35, negative in 4
E3, W1 Same wearer, sessions five days apart (band re-donned) 4 0.327 worse than constant 0.205
E3, W2 Same wearer, sessions 30 minutes apart (band re-donned) 4 0.160 0.202
E4 Across wearers (train on all of one wearer, test on the other) 4 0.272 28.5 deg per joint worse than constant 0.216 -0.96 to -0.35, all negative

E3 and E4 runs share training data and sessions, so they are not independent; the report gives ranges over runs, not confidence intervals.

1

Random splits flatter the model 2.4 to 2.6 times

E1 and E2 use identical data and the same model; only the split differs, and the error rises from 0.043 to 0.110. At 17 to 19 Hz, neighbouring frames are near-duplicates, so a random split leaks near-copies of test frames into training. The factor is specific to this sampling rate and these movements.

2

Without calibration, it is worse than the average pose

Five days and a re-donning apart, the model scores 0.327 against 0.205 for a constant that predicts the training-mean pose, in every one of four runs. Across wearers it scores 0.272 against 0.216. Thirty minutes apart it kept some skill (0.160 against 0.202).

3

Label-free standardization removes that failure

Z-scoring each cell with statistics from unlabeled data of the test session brings same-wearer error to 0.169 against 0.223 for the constant on the same frames, using only the first 30 s. Across wearers it stays near the constant.

Label-free standardization

Part of the change between sessions is a per-cell offset and scale shift. Each cell is z-scored with the mean and standard deviation of its own session, using no labels from the test session. The whole-session rows are an upper bound a live system would not have; the first-N-seconds rows use only the start of the test session for the statistics and score the frames after it. Raw model and constant are scored on the same frames.

Protocol Calibration data Runs Standardized Raw model Constant
E3, same wearer Whole session 8 0.147 0.244 0.203
E3, same wearer First 10 s 8 0.179 0.246 0.210
E3, same wearer First 30 s 8 0.169 0.250 0.223
E3, same wearer First 60 s 6 0.172 0.274 0.220
E3, W1 (five days) Whole session 4 0.160 0.327 0.205
E3, W1 (five days) First 30 s 4 0.190 0.327 0.209
E3, W2 (30 minutes) Whole session 4 0.134 0.160 0.202
E3, W2 (30 minutes) First 30 s 4 0.149 0.173 0.236
E4, cross-wearer Whole session 4 0.199 0.272 0.216
E4, cross-wearer First 10 s 4 0.217 0.275 0.218
E4, cross-wearer First 30 s 4 0.196 0.279 0.221
E4, cross-wearer First 60 s 4 0.187 0.284 0.218

At 60 s the 49 s session is skipped, so the same-wearer rows drop to 6 runs.

What it recovers: for the same wearer, standardization removes the worse-than-constant failure and recovers part of the within-session skill. With 30 s of unlabeled calibration data the error is 0.169 against 0.223 for the constant on the same frames, about midway between the constant and the within-session error. For W1, five days apart, it takes the error from 0.327 raw to 0.190.
What it does not recover: across wearers the error stays near the constant level (0.187 to 0.217 causal against constants of 0.218 to 0.221), and the standardized model explains none of the cross-wearer variance. Standardization removes an offset and a scale per cell; it cannot undo a change in which cells respond to which finger. The same-wearer against cross-wearer ordering rests on two wearers and is an observation, not a finding.
The model: one random forest (100 trees) per hand per condition, mapping the 60 raw cell values of a single frame to 15 wrist-relative finger targets. No temporal context. It is a floor, chosen so that differences between protocols cannot be blamed on model capacity. Temporal models and supervised per-session calibration are the obvious next experiments. A one-band LASK5 subset on the legacy V3 band shows the same pattern for piston position.
Limits: two wearers, four sessions, one hardware revision (FlexGrid V4) for the main subset. VR hand tracking is itself an estimator, so label noise is reduced but not eliminated. Flexion is normalized with nominal joint ranges rather than per-person calibration. Band placement was not measured.
Participants and ethics: two wearers, the author (W1) and an adult member of his family (W2), both with intact hands. Both took part voluntarily and agreed to the pseudonymised release of their recordings. No institutional ethics review was sought, because the work is independent and self-funded. Nothing here says how the system behaves for people with limb difference.
What changed in v0.3: the report has a new title and adds the constant baseline on every run, label-free standardization, blocked cross-validation and robustness checks, and reproduction from the public release. Withdrawn: v0.2 claimed that re-donning the band rivals changing the wearer, because the five-day cross-session error was as high as or higher than the cross-wearer error. Both raw numbers are worse than the constant baseline, so neither ordering of two failed models says anything about donning or wearer identity; after standardization the ordering reverses. The version note in the report has the details.

Reproduce it

Every table and figure in the main text is produced from the public release by the scripts in analysis/paper/. You need Python 3.12 and git. About 50 MB is downloaded anonymously, and the whole procedure took about 12 minutes in a clean-room run. These are the Windows commands; REPRODUCE.md on GitHub also gives the Linux and macOS form.

git clone https://github.com/Open-Muscle/OpenMuscle-Software.git
cd OpenMuscle-Software
py -3.12 -m venv .venv-paper
.venv-paper\Scripts\python -m pip install -r analysis/paper/requirements.txt
.venv-paper\Scripts\python analysis/paper/analyze.py --source hf --out results_check.json
.venv-paper\Scripts\python analysis/paper/analyze_lask5.py --source hf --out results_lask5_check.json
.venv-paper\Scripts\python analysis/paper/paper_numbers.py
.venv-paper\Scripts\python analysis/paper/figures.py

The forests use a fixed seed and the package versions are pinned, so the same versions give the same numbers. Appendix B of the report uses a capture that is not in the release, and part of Appendix A comes from a committed May 2026 run; the report says so.

How to cite

If you use the data, please cite the dataset by its DOI. If you use the results or the hardware designs, cite the paper as well. Either way, credit the OpenMuscle project, as CC BY 4.0 asks. The paper DOI is pending and will be added here once it is issued.

Plain text (paper)

Moghadam, T. (2026). OpenMuscle: An Open Forearm Force-Myography Dataset with Continuous Finger Labels, and Baselines Across Sessions and Wearers. OpenMuscle technical report v0.3. https://openmuscle.org/research (DOI pending)

BibTeX (paper and dataset)

@misc{moghadam2026openmuscle,
  author = {Moghadam, Tory},
  title = {{OpenMuscle: An Open Forearm Force-Myography Dataset with Continuous Finger Labels, and Baselines Across Sessions and Wearers}},
  year = {2026},
  url = {https://openmuscle.org/research},
  note = {OpenMuscle technical report v0.3, DOI pending},
}

@misc{moghadam2026openmusclefmg,
  author = {Moghadam, Tory},
  title = {{OpenMuscle FMG dataset}},
  year = {2026},
  publisher = {Hugging Face},
  doi = {10.57967/hf/11416},
  url = {https://huggingface.co/datasets/openmuscle/openmuscle-fmg},
}

Author: Tory Moghadam, ORCID 0009-0001-8126-247X.