openmuscle/openmuscle-fmg

OpenMuscle FMG: forearm pressure grid with finger labels

Forearm force-myography (FMG) recordings from the open-hardware OpenMuscle FlexGrid band: a grid of 60 pressure cells worn on the forearm, recorded together with finger-movement labels from the LASK5 spring-loaded four-piston finger labeler or from Meta Quest 3S hand tracking. The dataset is for researchers in human-machine interfaces, prosthetic sensors and VR interfaces. The targets are continuous finger flexion and extension, so the task is per-frame regression (predict how far each finger is curled from forearm pressure), not gesture classification.

Release 0.1 (November 2026). Every count, duration and metric on this card was computed from the packaged files, the July 2026 data audit, or the released results of the OpenMuscle technical report v0.2.

Config	Captures	Rows	Minutes	Parquet (MB)	Labels	Board (from metadata)	Wearers	Recorded (UTC)
lask5_labeled (default)	23	43,373	28.6	3.53	LASK5, 4 pistons	FlexGrid V3 (20), unrecorded pre-V4 (3)	P01	2026-05-14 to 2026-05-31
quest_labeled	25	41,051	24.6	58.81	Quest hand tracking, 25 joints	FlexGrid V4 (14), FlexGrid V3 (11)	P01, P02	2026-05-25 to 2026-07-03
legacy_2025	13	13,373	27.9	1.42	LASK5, 4 pistons, partial coverage	2025 prototype, 64 cells	P01	2025-07-16 to 2025-07-17
supplementary_unlabeled	14	250,581	522.8	3.48	none	2025 prototype (13), unrecorded pre-V4 (1)	P01	2025-07-07 to 2026-05-14
Total	75	348,378	603.9	67.24

The two supervised configs that pass the audit's publish tier (lask5_labeled and quest_labeled) hold 48 captures, 84,424 rows and 53.2 minutes. Rows are sensor frames; a two-band capture interleaves the rows of both bands.

Important: no capture in this release pairs the FlexGrid V4 band with the LASK5. All LASK5-labeled data was recorded in May 2026, before V4 existed (see Known caveats). The V4 data is the 14 V4 captures inside quest_labeled; select them with device == "flexgrid-v4".

Who this is for

• Prosthetics control research: train and benchmark models that map forearm pressure to finger flexion, with the temporal, cross-session and cross-wearer baselines below to compare against.
• VR and HMI input: camera-free, glove-free finger input; the Meta Quest 3S hand-tracking labels (25 joints per frame) pair with every band frame in quest_labeled.
• Wearable sensing methods: donning variance, cross-board and cross-wearer transfer, evaluation protocol for low-cost sensor arrays, and long-wear streams (supplementary_unlabeled, 522.8 minutes) for drift and self-supervised studies.

What it is not. This is not a gesture-class dataset (the targets are continuous finger positions, there are no gesture labels), not clinical data (the records hold no health fields), and it has two wearers only (P01 and P02), so cross-wearer results are directional.

Which board recorded each capture

The FlexGrid V4 board was fabricated in June 2026, so anything earlier is V3 or older. The device column and manifest.csv (device, band_ids, board_evidence) label each capture only by what its metadata records:

device	Captures	Configs	Recorded (UTC)	Evidence
flexgrid-v4	14	quest_labeled	2026-06-26 to 2026-07-03	.meta.json names unit flexgrid-d7af0b (11) or flexgrid-1524c3 (3); session records show firmware v4.0.0 for both units
flexgrid-v3	31	lask5_labeled (20), quest_labeled (11)	2026-05-16 to 2026-06-05	.meta.json names unit flexgrid-v3-02
unrecorded-pre-v4	4	lask5_labeled (3), supplementary_unlabeled (1)	2026-05-14 to 2026-05-15	No board in the metadata (the sensor stream id flexgrid-legacy names a wire format, not a board); recorded before V4 existed
flexgrid-prototype-2025	26	legacy_2025 (13), supplementary_unlabeled (13)	2025-07-07 to 2025-07-17	No metadata; recorded before the V3 board existed

Quick start

from datasets import load_dataset

lask5 = load_dataset("openmuscle/openmuscle-fmg", "lask5_labeled", split="train").to_pandas()
quest = load_dataset("openmuscle/openmuscle-fmg", "quest_labeled", split="train").to_pandas()

cells = [f"R{r}C{c}" for r in range(4) for c in range(15)]          # 60 features
pistons = ["label_0", "label_1", "label_2", "label_3"]               # LASK5 targets
v4 = quest[quest["device"] == "flexgrid-v4"]
take = v4[v4["capture_id"] == "capture_1783114592"]                  # a two-band capture
left_band = take[take["arm"] == "left"].sort_values("ts_hub_ms")

Each parquet file is one capture, so pandas.read_parquet("data/lask5_labeled/capture_1778986907.parquet") works on a local clone too. manifest.csv lists every capture with its config, wearer, session, rows, duration, board and source path.

What was recorded

Band (features). The V3 and V4 FlexGrid bands carry a 4 x 15 grid of Velostat pressure cells (60 channels) on a flexible PCB, scanned by an ESP32-S3 through an analog multiplexer and read as raw 12-bit ADC counts (0 to 4095). The band streams JSON frames over Wi-Fi UDP to a PC hub, which timestamps every frame on its own clock.

• FlexGrid V4 (fabricated June 2026), 14 captures: adds a 6-axis IMU (ICM-42688 family, scale in each capture's .meta.json). Two physical units, flexgrid-d7af0b and flexgrid-1524c3 (left and right by the project's convention when both are worn). Measured 17.0 to 19.2 Hz per band. Nine captures record both bands at once. Eight of them are the report's analysis set (the paper_session column of manifest.csv). The ninth, capture_1783111690 (P02, 27.5 s, taken earlier in the same donning as the report's W2-s1 capture), passed the audit and is published here, but the report does not analyse it. The report's two named exclusions, a 9.7 s single-band fragment and a capture with a 29.5 s tracking-loss gap, are not in this release (see Known caveats and Excluded captures).
• FlexGrid V3 (validated 2026-05-13), 31 captures, unit flexgrid-v3-02: no IMU columns. Measured 13.7 to 24.8 Hz per capture (median 24.4 Hz).
• Unrecorded pre-V4 board, 4 captures of 2026-05-14 and 2026-05-15: 60 cells, no board id in the metadata, 31.4 to 33.2 Hz.
• 2025 prototype, 26 captures: an early FlexGrid board that predates V3. Most files carry a 4 x 16 grid (64 cells, R0C0 to R3C15); some supplementary files carry 4 x 15. Rows arrive at 2.7 to 9.8 Hz (median 9.0 Hz). Saturated cells (4095) and dead cells occur.

Labels (targets).

• LASK5 (labeler = lask5). A handheld labeler with four spring-loaded, hall-sensor finger pistons and a thumb joystick, held in the hand of the banded arm. The CSV kept the four piston channels as label_0 to label_3, normalized to [0, 1] in lask5_labeled; pairing window 100 ms. The finger order of the four channels is not confirmed in the project documentation.
• Meta Quest 3S hand tracking (labeler = quest_hand). A WebXR page on the headset streams 25 hand joints per frame (position plus quaternion, 7 floats per joint, 175 values) to the hub, where each band frame is paired with the nearest label frame within a 175 ms window. Raw values are in label_0 to label_174; the joint and channel of each column is in that capture's .labels.schema.json. Seven V4 captures (2026-07-03) also carry the 15 canonical wrist-relative labels (lbl_flex_*, lbl_ang_*) written at capture time.

Why the LASK5 subset is listed first: a hypothesis, not a result. The author's rationale is that the LASK5 pistons load the fingers with springs, so the forearm muscles contract against resistance and the pressure signal on the band should be more pronounced, while camera and VR labels capture free-air finger curls with less contraction. The technical report v0.2 tests this once, in section 5.6, by comparing the band signal itself on the V3 band, which recorded under both labelers (the two labelers' R^2 values cannot settle it, because their targets differ). Over all captures the report's 20 V3 LASK5 recordings vary 1.72x more per cell than its 10 V3 Quest recordings (per-cell standard deviation 26.8 against 15.6 counts, p = 0.003). Against the late-May LASK5 sessions nearest in date to the Quest recordings the ratio is 1.12x (p = 0.16). The report therefore concludes that the data neither confirm nor rule out the hypothesis: any labeler effect is smaller than the session-to-session variation of a single band, and the two labelers were used for different movement routines. No V4 recording has LASK5 labels, so the test is V3 only. The report's other LASK5 results (the ladder in section 5.4 and the single-session benchmark in section 5.5, below) are also on the V3 board, with different targets and metrics from the Quest baselines. Section 5.5 reports that the piston labels sit at the fully-closed rail in 64 to 79% of frames. A direct test, which this release cannot provide, would alternate LASK5 and free-air blocks within one session on a V4 band left in place.

The collection protocol is on one page in PROTOCOL.md; the column schema is in DATA_SCHEMA.md.

Columns

Every config starts with six provenance columns added for this release:

Column	Type	Meaning
capture_id	string	Capture name; the number is the Unix start time in seconds
wearer	string	Pseudonymous wearer code, P01 or P02
session_id	string, nullable	Recorded session (one donning), named by its local start time; null when no session was recorded
device	string	Board from metadata: flexgrid-v4, flexgrid-v3, unrecorded-pre-v4 or flexgrid-prototype-2025
labeler	string	lask5, quest_hand or none
arm	string	Forearm the band was worn on, left or right (derived, see Known caveats)

lask5_labeled (72 columns)

Columns	Type	Meaning
timestamp	float64	Hub receive time, Unix seconds
device_id	string	Unit id as recorded (flexgrid-v3-02, or flexgrid-legacy for the 3 unrecorded-board captures)
R0C0 to R3C14	int16	60 pressure cells, raw ADC counts, row-major (4 rows x 15 columns)
label_0 to label_3	float64	LASK5 piston channels, [0, 1]; every row has all four

quest_labeled (274 columns)

Columns	Type	Present in	Meaning
timestamp	float64	11 V3 captures	Hub receive time, Unix seconds
ts_hub_ms	int64	14 V4 captures	Hub receive time, Unix milliseconds (the one sync clock)
role	string	14 V4 captures	Band role as recorded, left or right
device_id	string	all	Physical unit id
R0C0 to R3C14	int16	all	60 pressure cells, raw ADC counts, row-major (4 rows x 15 columns)
label_0 to label_174	float64	all	Raw Quest joints: 25 joints x [px, py, pz, rx, ry, rz, rw], headset world frame
imu_gx imu_gy imu_gz imu_ax imu_ay imu_az	int32	14 V4 captures	Band IMU raw counts (gyro, accel); multiply by imu_scale in the meta sidecar
lbl_imu_*	float64	none	Labeler-device IMU slot; empty in this release (Quest has no IMU stream)
forearm_roll_deg, palm_up	float64, int8	12 V4 captures	Forearm roll in degrees and a palm-up flag, derived from the matched Quest hand (label side, not the band)
lbl_flex_thumb to lbl_flex_pinky	float64	7 V4 captures	Tier-1 canonical label: per-finger flexion in [0, 1], 0 = extended, 1 = curled
lbl_ang_*_mcp, lbl_ang_*_pip, lbl_ang_thumb_ip	float64	7 V4 captures	Tier-2 canonical label: per-joint flexion in degrees, flexion-positive

Columns a capture does not have are null in its rows, so every file in a config shares one schema. The canonical labels can be re-derived for the other 18 captures from label_* with openmuscle.hand_angles in OpenMuscle-Software, as the baseline report did. All 41,051 rows carry a complete 175-value label vector. The 14 V4 captures hold 29,082 rows (15.8 minutes); the 11 V3 captures hold 11,969 rows (8.9 minutes).

legacy_2025 (75 columns)

timestamp, R0C0 to R3C15 (int16, 64 cells, raw ADC counts), label_0 to label_3 (float64, LASK5 pistons). Label coverage per capture runs from 0.263 to 1.0; 8,719 of 13,373 rows (65.2%) have all four labels. The first seven captures store the pistons in raw units (maximum 1,961.9); the last six store them normalized to [0, 1]. Four captures contain cells at full scale (4095) in 292 rows.

supplementary_unlabeled (71 columns)

timestamp and R0C0 to R3C15 (int16). Files recorded with a 60-cell grid leave R0C15 to R3C15 null. The longest capture runs 18,054.3 s (5.0 h); eight captures contain full-scale (4095) cells in 1,357 rows. These long-wear streams suit drift, wear, noise and self-supervised studies, not supervised finger regression.

Sidecars next to each parquet: <capture_id>.meta.json (unit ids, firmware reset cause and battery, IMU scale, band roles, match window, session id; 47 captures) and <capture_id>.labels.schema.json (Quest label layout; 25 captures).

How to split the data

Do not use random row splits for evaluation. At these frame rates adjacent frames are near-duplicates, so a random split puts near-copies of test frames in training. On the report's eight two-hand V4 captures the same random forest scores a finger-flexion MAE of 0.043 under a random 80/20 split and 0.110 under a temporal 80/20 split, about 2.6x worse. On the LASK5 session the best R^2 drops from 0.969 (shuffled) to 0.702 (time split).

Recommended, in increasing difficulty:

1. Temporal within capture or session: train on the first part of the timeline, test on the last part.
2. Leave one session out: group by session_id (one donning). Fifteen lask5_labeled captures, nine quest_labeled captures and all 2025 captures have no recorded session; treat each of those captures as its own group, or group by date_utc in manifest.csv, knowing that one date can hold more than one donning.
3. Leave one wearer out: group by wearer (only quest_labeled has two).

Keep boards apart or treat device and device_id as covariates: in unpublished project analyses, models trained on one physical V4 unit transferred poorly to the other.

Baselines

LASK5, FlexGrid V3

From the OpenMuscle technical report v0.2, section 5.5 (numbers in analysis/anniversary/results/bench.md of OpenMuscle-Software). Session 2026-05-16T21-58-59, wearer P01, four captures (capture_1778986907, capture_1778987104, capture_1778987285, capture_1778987468), 9,169 rows; targets are the four piston channels.

Model, features	R^2, last 20% of each recording held out	R^2, shuffled 80/20
Ridge, raw + differences + 5-frame rolling mean (180)	0.702	0.794
Extra trees (200), same 180 features	0.697	0.969
Random forest (100), raw cells (60)	0.587	0.915

Quest hand tracking, FlexGrid V4

From the report's results.json. One 100-tree random forest per hand maps the 60 raw cells of a single frame to the 15 canonical targets (re-derived from the raw Quest labels). Data: eight two-hand V4 captures in quest_labeled, 22,291 frames. The report's wearers W1 and W2 are P01 and P02 here.

Report session	Wearer	Captures
W1-s1	P01	capture_1782662414, capture_1782662954, capture_1782663139
W2-s1	P02	capture_1783111763
W2-s2	P02	capture_1783113626, capture_1783113806
W1-s2	P01	capture_1783114480, capture_1783114592

The paper_session column of manifest.csv carries the same mapping.

Protocol	Flexion MAE (fraction of full curl)	Angular MAE, mean (range)	Flexion R^2 range
E1 within session, random 80/20	0.043	4.9 deg (4.1 to 6.0)	0.84 to 0.91
E2 within session, temporal 80/20	0.110	12.7 deg (7.8 to 18.7)	-2.37 to 0.77 (0.21 to 0.77 in six of eight runs; negative for both hands of the 49 s W2-s1 session)
E3 cross-session, same wearer	0.24 (P02 0.16, P01 0.33)	25.7 deg (16.0 to 42.0)	-2.34 to 0.35
E4 cross-wearer	0.27	28.5 deg (27.1 to 31.9)	-0.96 to -0.35

Donning variance rivals wearer identity: for P01, a second session five days later with the band re-donned (flexion MAE 0.25 to 0.42) did at least as badly as transferring across wearers (0.26 to 0.31).

The two baseline tables use different boards, targets and metrics and must not be compared with each other.

Known caveats

• No V4 LASK5 capture. The only LASK5 session on a V4 band (2026-07-16, unit flexgrid-1524c3) produced two captures, capture_1784216647 (37.0 s) and capture_1784216703 (4.8 s), with zero paired rows: their metadata reports a match rate of 0.0 and their label stream is empty, so the CSVs hold only a header. They were recorded after the July 2026 audit, so no audit verdict covers them, and there is no data to package. No audit verdict excluded any LASK5 V4 capture.
• Two wearers. P01 contributed every config; P02 contributed 5 quest_labeled V4 captures (12,912 rows, 6.3 minutes), all on 2026-07-03. Cross-wearer conclusions are directional.
• Donning variance. Band placement was not measured or controlled between sessions. Expect large shifts between sessions and between the two physical units.
• Arm is derived, not recorded. Per the maintainer's records, single-band captures before 2026-06-30 were worn on the right forearm even where role says left (a default before bands were assigned roles). From 2026-06-30, and in every two-band capture, role is physical (flexgrid-d7af0b left, flexgrid-1524c3 right) and arm follows it. The 2025 captures are attributed to the right forearm from the same records.
• Hand tag of four V4 captures. The single-band V4 captures of 2026-06-26 to 2026-06-28 (right forearm) carry a left-hand Quest tag. Within-session models predict their labels well in project analyses, but which hand the headset tracked is not recorded; take care when mixing them into cross-arm work. In two-band captures each band's rows carry the labels of the hand on the same arm, although .labels.schema.json names only the first hand seen.
• Excluded capture. capture_1783113997 (P02, V4, session 2026-07-03T16-18-43) passed the automatic checks but has a 29.5 s Quest tracking-loss label gap, 23% of the take, and is not included. The automatic verdicts do not detect label freezes, so other short tracking losses may remain.
• Dead column. On the left V4 band during W1-s1 (2026-06-28), sensor column C3 (four cells) read zero for most of the session (report v0.2 section 4.1).
• LASK5 labels. The four channels are piston positions, not joint angles, and the finger order is unconfirmed. The pistons rest at the fully-closed rail in most frames (report v0.2 section 5.5).
• Saturation and dead cells in the 2025 data. See the counts above; the audit flagged heavy ADC saturation in several supplementary captures.
• Labels are estimates. Headset hand tracking has its own error, not measured here. forearm_roll_deg and palm_up come from the Quest, so they are labels, not inputs available at inference. lbl_imu_* is empty.
• Raw streams not included. Per-device .sensor.jsonl and .label.jsonl streams (including the LASK5 thumb joystick and per-joint Quest validity flags) are not part of this release.

Pseudonymization and privacy

Wearers are identified only as P01 and P02, assigned in order of first recording; captures without a session record are attributed to P01 per the maintainer's records. The name-to-code key is not published. Free-text fields (session names, notes, gesture notes, subject fields) were removed from every sidecar; only machine provenance remains. The source records hold no age, sex, handedness or health fields. The release contains forearm pressure, band IMU and hand-joint trajectories, no images, audio or video. Every text file and every parquet string column in the package was scanned for names, email addresses and place names before release.

Relationship to the GitHub drop

The first public drop, OpenMuscle-Dataset (CC BY 4.0, raw CSV), holds the 2025 captures. This release supersedes it for the FlexGrid captures and adds the 2026 recordings:

• GitHub merged/ (18 files): 13 are legacy_2025 here; the audit excluded the other 5 (2 with no data rows, 1 flat dud, 2 micro-takes).
• GitHub flexgrid/ (16 files): 13 are in supplementary_unlabeled; the audit excluded the other 3 (2 with no data rows, 1 micro-take).
• GitHub curated/ and legacy/ (pre-FlexGrid text captures) were not audited and stay on GitHub only.

The GitHub README labels the 2025 folders "FlexGrid V3"; the hardware revision history dates V3 to May 2026, so this card calls the 2025 board a prototype.

Excluded captures

The July 2026 audit classified 124 capture files (66 from 2026 in its 2026-07-03 re-scan, 58 from 2025); 75 are packaged. The other 49 are its 48 EXCLUDE verdicts (files with no data rows, micro-takes under 10 s or 200 rows, flat duds where most cells never move) and capture_1783113997 (label gap, above). Captures recorded after the audit (2026-07-16) and simulated or smoke-test files are not included either.

License

CC BY 4.0. Use, share and adapt for any purpose, including commercially, with attribution. Full text in LICENSE. Software is MIT (OpenMuscle-Software) and hardware is CERN-OHL-S-2.0 (OpenMuscle-FlexGrid); see those repositories.

Citation

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

@misc{moghadam2026openmusclefmg,
  author       = {Moghadam, Tory},
  title        = {{OpenMuscle FMG}: Forearm Pressure Grid Recordings with Finger Labels},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/openmuscle/openmuscle-fmg}},
  note         = {Release 0.1. ORCID 0009-0001-8126-247X. DOI: 10.57967/hf/11416}
}

@misc{moghadam2026openmuscle,
  author       = {Moghadam, Tory},
  title        = {OpenMuscle: Forearm Force-Myography with VR Hand-Tracking Ground Truth: Dataset and Honest Baselines},
  year         = {2026},
  howpublished = {OpenMuscle technical report v0.2},
  url          = {https://openmuscle.org/research},
  note         = {DOI: pending}
}

Links

• Research page: https://openmuscle.org/research
• Dataset, first drop: https://github.com/Open-Muscle/OpenMuscle-Dataset
• Capture, training and analysis software: https://github.com/Open-Muscle/OpenMuscle-Software
• FlexGrid hardware: https://github.com/Open-Muscle/OpenMuscle-FlexGrid
• Band firmware: https://github.com/Open-Muscle/FlexGridV4-Firmware and https://github.com/Open-Muscle/FlexGridV3-Firmware
• LASK5 labeler: https://github.com/Open-Muscle/OpenMuscle-LASK5
• Project hub: https://github.com/Open-Muscle/OpenMuscle-Hub