FEEL-Benchmark

A multimodal egocentric manipulation dataset combining Project Aria smart-glasses RGB video and MPS hand tracking with synchronized PX6AX-GEN3-DP-S2015-Elite fingertip tactile-sensor readings. Released in LeRobot v3.0 format (GR00T path variant).

Overview

• 52 trials, 3,317 episodes, ~2.8 h of segmented manipulation, ~4.6 GB
• Recording dates: 2026-01-10, 2026-01-21, 2026-01-28, 2026-04-30, 2026-05-06, 2026-08-05, 2026-08-06, 2026-08-07
• Camera: Project Aria RGB (1408×1408, 10 fps, MPEG-4 Part 2 / mpeg4, yuv420p)
• Tactile: two PX6AX-GEN3-DP-S2015-Elite fingertip sensor modules on the right hand; per module and per frame, a 3-axis force vector (Fx, Fy, Fz) plus its magnitude, resampled to the 10 fps video grid (raw logs ran at ~900 Hz in 2026-01/04/05 and ~91 Hz in the August sessions)
• Hand pose: right-hand thumb-tip and index-tip positions from Aria MPS hand tracking, plus a per-frame validity flag
• Mean episode length: 3.1 s (101,219 frames @ 10 fps)
• 4 collectors, anonymized as Collector A / B / C / D

Release history

	v1 (2026-05-07)	v2 (2026-08-11)	Δ	Change
Trials	26	52	+26	+100%
Episodes	700	3,317	+2,617	+374%
Frames	39,497	101,219	+61,722	+156%
Duration (min)	65.8	168.7	+102.9	+156%
Size	1.99 GB	4.62 GB	+2.62 GB	+132%
Collectors	1	4	+3	+300%
Mean episode length	5.6 s	3.1 s	−2.5 s	−45%

v2 adds the 2026-08-05/06/07 sessions (+2,637 episodes) and replaces P25_bottle_collectorC and P75_bottle_collectorC with re-segmented versions (−20 episodes). All v1 trials were recorded by a single person (Collector C, who did not record again after 2026-05-06); the August sessions were run by three different people (Collectors A, B and D), which makes cross-collector evaluation possible. The shorter mean episode length reflects finer segmentation in the v2 sessions (2026-08-07 averages 2.3 s/episode vs. 5.6 s in v1), not a change of task.

v2.1 (2026-08-11) adds the per-axis tactile columns action.tactile.module1.force_xyz and action.tactile.module2.force_xyz to every episode parquet, recovered from the raw tactile logs and resampled with the same interpolation that produced the existing force_total columns. Trial, episode and frame counts are unchanged, as are the 14 pre-existing columns and all videos.

It also renames the 26 Collector C trial directories to carry a collectorC infix, so that all four collectors are identified the same way instead of Collector C being the implicit "no infix" case. The object name and trailing index are unchanged — 20260110/Jelly_1_lerobot is now 20260110/Jelly_collectorC_1_lerobot, 20260506/Wz_lerobot is now 20260506/Wz_collectorC_lerobot. File contents were moved unmodified (identical LFS hashes), but paths recorded before 2026-08-11 need updating.

Dataset statistics

By recording date

Date	Trials	Episodes	Frames	Duration (min)
2026-01-10	3	42	3,565	5.9
2026-01-21	7	137	11,106	18.5
2026-01-28	5	132	8,481	14.1
2026-04-30	2	82	4,761	7.9
2026-05-06	9	287	11,584	19.3
2026-08-05	1	104	2,842	4.7
2026-08-06	6	302	8,193	13.7
2026-08-07	19	2,231	50,687	84.5
Total	52	3,317	101,219	168.7

By object category

Category	Trials	Episodes	Frames	Duration (min)
Bottle	19	1,534	43,040	71.7
Jelly	8	330	16,123	26.9
Sponge	6	436	13,594	22.7
Water	6	398	11,418	19.0
Wangzai	7	341	10,025	16.7
Chip	4	259	5,507	9.2
Coke	2	19	1,512	2.5

By collector

Four people recorded this dataset. Each is identified by a collectorA / collectorB / collectorC / collectorD infix in the trial directory name:

Collector	How to identify	Trials	Episodes	Frames	Duration (min)
Collector A	collectorA infix	10	1,036	27,271	45.5
Collector B	collectorB infix	5	829	15,130	25.2
Collector C	collectorC infix	26	680	39,497	65.8
Collector D	collectorD infix	11	772	19,321	32.2

Collector C recorded the whole 2026-01/04/05 portion alone and did not record again afterwards; the three August sessions were run by Collectors A, B and D. The mapping between these labels and the people who recorded the data is not released.

The 2026-08-07 session is the only one where several collectors (A, B and D) recorded the same objects, which makes it usable as a cross-collector generalization split.

Layout

Datasets are grouped by recording session date:

FeelAuthors/FEEL-Benchmark/
├── 20260110/  Jelly_collectorC_{1,2,3}                                                    (3 trials)
├── 20260121/  Half_coke_collectorC_2, Half_water_collectorC_1, Jelly_collectorC_{5,6,7},
│              Quarter_coke_collectorC_0, Quarter_water_collectorC_2                       (7 trials)
├── 20260128/  Half_water_collectorC_{3,4}, Sponge_collectorC_{1,2,3}                      (5 trials)
├── 20260430/  Empty_bottle_collectorC_1, Full_bottle_collectorC_1                         (2 trials)
├── 20260506/  Chip_{intact,crushed}_collectorC, Empty_Want_Milk_collectorC,
│              Empty_wz_collectorC, P25_bottle_collectorC, P50_bottle_collectorC,
│              P75_bottle_collectorC, Want_Milk_collectorC, Wz_collectorC                  (9 trials)
├── 20260805/  sponge_collectorD_2                                                        (1 trial)
├── 20260806/  bottle_p0_collectorD_{1,2}, bottle_p50_collectorD_{1,3},
│              empty_want_milk_collectorD_2, jelly_collectorD_2                           (6 trials)
└── 20260807/  bottle_p0_collectorA_1, bottle_p100_collectorA_2,
               bottle_p100_collectorB_2, bottle_p100_collectorD_1,
               bottle_p50_collectorA_1, bottle_p50_collectorB_1, bottle_p50_collectorD_1,
               bottle_p75_collectorA_1, bottle_p75_collectorB_1, bottle_p75_collectorD_1,
               chip_collectorA_1, chip_collectorD_1,
               empty_want_milk_collectorA_{1,2}, jelly_collectorA_1,
               sponge_collectorA_1, sponge_collectorB_1,
               water_p25_collectorA_1, water_p25_collectorB_1                            (19 trials)

p0 / p25 / p50 / p75 / p100 denote the fill level of the container.

Each *_lerobot/ directory follows LeRobot v3.0:

<trial>_lerobot/
├── meta/
│   ├── info.json          # feature schema, fps, episode counts
│   ├── episodes.jsonl     # per-episode metadata
│   ├── episodes/chunk-000/file-000.parquet
│   ├── modality.json      # GR00T state/action field map
│   ├── stats.json
│   ├── tasks.jsonl
│   └── tasks.parquet
├── data/chunk-000/episode_NNNNNN.parquet
└── videos/chunk-000/observation.images.main/episode_NNNNNN.mp4

Feature schema

Each row in episode_*.parquet:

Field	dtype	shape	Notes
index	int64	(1,)	Global frame index
episode_index / episode_id	int64/int32	(1,)
episode_chunk	int64	(1,)	Chunk index (always 0)
frame_index / step	int64	(1,)	Per-episode frame index
task_index	int64	(1,)	Index into tasks.jsonl
timestamp / timestamp_ns	float64/int64	(1,)	Aria device clock
observation.state	float32	(7,)	Thumb tip xyz (3) + index tip xyz (3) + pinch distance (1)
action	float32	(2,)	[module1_force_total, module2_force_total]
status.hands.valid_right	bool	(1,)	Right-hand tracking validity
action.tactile.left.force_total	float32	(1,)	Sensor module 1 force magnitude (raw counts)
action.tactile.right.force_total	float32	(1,)	Sensor module 2 force magnitude (raw counts)
action.tactile.module1.force_xyz	float32	(3,)	Sensor module 1 3-axis force [Fx, Fy, Fz] (raw counts)
action.tactile.module2.force_xyz	float32	(3,)	Sensor module 2 3-axis force [Fx, Fy, Fz] (raw counts)
observation.images.main	video	1408×1408×3	Aria RGB stream (mpeg4, yuv420p, 10 fps)

Note on the left / right tactile column names

action.tactile.left.force_total and action.tactile.right.force_total are not left-hand and right-hand sensors. They are the two sensor modules of one instrumented right hand; the left/right naming is a legacy artifact of the converter, which assigns modules by column order in the raw log rather than by name.

Which fingers those two modules sat on is not the same in every session:

Trials	Module 1 (…left.force_total, …module1.force_xyz)	Module 2 (…right.force_total, …module2.force_xyz)
the 45 trials under 20260110/ … 20260506/ and 20260807/	thumb tip	index fingertip
the 7 trials under 20260805/ and 20260806/	middle-finger tip	little-finger tip

Treat the columns as "module 1 / module 2" and consult this table before pooling sessions. modality.json exposes the four values as module1_force_total / module2_force_total / module1_force_xyz / module2_force_xyz, and each entry's description field names the finger that module was mounted on in that session.

The force_total value is the magnitude sqrt(Fx² + Fy² + Fz²) of that module's 3-axis combined force, in raw signed-16-bit sensor counts — not newtons. No force calibration was applied, so the values are suitable for relative/temporal analysis (contact onset, grip modulation) but not for absolute force comparison across sessions. The distributed per-taxel readings are not part of this release.

The two force_xyz columns use the neutral module1 / module2 names rather than the legacy left / right; module1.force_xyz pairs with left.force_total, and module2.force_xyz with right.force_total.

Note on the 3-axis force components

action.tactile.module{1,2}.force_xyz hold the per-axis combined force [Fx, Fy, Fz] in the same raw signed-16-bit counts as force_total. Two caveats:

• The axes are the sensor module's own local frame. Each module's orientation on the fingertip was not recorded, so the axes are not expressed in a hand, camera or world frame, and they are not directly comparable across modules or sessions. Empirically Fz carries most of the magnitude (mean |Fz|/‖F‖ = 0.78; Fz is the largest-magnitude axis in 67–70 % of frames), consistent with Fz being the axis normal to the sensing pad and Fx/Fy the two shear components — but that is an observation about this data, not a calibrated fact.
• ‖(Fx, Fy, Fz)‖ is not exactly equal to force_total. The magnitude and the vector were each interpolated onto the video frame grid independently, and linear interpolation does not commute with taking a norm when the force direction rotates between two raw samples. The two agree exactly at raw sample times; between them they differ by ≤ 0.01 counts on 96.8 % of frames, with a p99 of 0.03 and a worst case of 0.86 counts (force_total itself has p50 = 6.1, p99 = 79.9, max = 143.4 counts). Use force_total when you want the scalar the original pipeline produced, and force_xyz when you want direction.

Across the 69.1 M raw samples behind this release the per-axis values span [-128, 255] counts, and [-128, 136] after resampling; the extreme values are each ~0.002 % of samples, so no material clipping occurs.

Note on tactile sampling

The tactile stream is logged independently of the camera and then linearly interpolated onto each video frame's timestamp, so the dataset carries exactly one magnitude and one 3-axis vector per module per frame at 10 Hz. Raw log rates differ by session:

Sessions	Trials	Raw log rate	Host-timestamp resolution
2026-01-10 … 2026-05-06	26	~897–901 Hz (1 ms period)	1 ms
2026-08-05, 2026-08-06	7	~91 Hz (11 ms period)	11 ms
2026-08-07	19	~91 Hz	50 ms — samples arrive in bursts of ~4 sharing one host timestamp

For the 2026-08-07 session the driver stamped each polled batch with a single host time, so the effective time resolution of that session's tactile stream is 50 ms even though rows arrive at ~91 Hz. Interpolation onto the 10 fps grid is therefore coarser there than the row count suggests.

observation.state holds the right-hand thumb tip (MPS hand landmark 4) and index fingertip (landmark 8) in the Aria world frame, followed by the Euclidean pinch distance between them. Positions are NaN on frames where the hand is not tracked — check status.hands.valid_right.

Note on video encoding

Videos are encoded as MPEG-4 Part 2 (mpeg4), not H.264 — the converter requested H.264 but the OpenCV build silently fell back to the mp4v encoder. Every decoder identifies the codec from the container, so playback and frame indexing are unaffected; meta/info.json reports "video.codec": "mpeg4" to match the actual streams.

Path format note

Each meta/info.json declares "path_format": "groot". The data_path and video_path templates match standard LeRobot v3.0 (data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet), so the dataset is consumable by both stock LeRobot tooling and NVIDIA Isaac GR00T loaders.

Sensors

• Project Aria (RGB, eye-tracking cameras, SLAM cameras, IMUs) with MPS post-processing for hand tracking. The recordings also contain eye gaze, SLAM trajectory and IMU streams, but this release exports only the RGB video and the two hand-landmark positions listed above — gaze, full 21-joint hand landmarks, left-hand data, device pose and IMU are not included in the parquet files.
• PX6AX-GEN3-DP-S2015-Elite fingertip tactile sensors — two modules on the right hand. The sensor hardware reports 3-axis combined force (each axis a 16-bit signed value) plus distributed per-taxel forces over a serial link; the logger recorded the two enabled modules at ~900 Hz (2026-01/04/05) or ~91 Hz (August). The combined force of each module is exported here both as a magnitude and as its three axis components, linearly interpolated onto the 10 fps video frame grid — the per-taxel data is not included in this release. See the note above for which fingers the two modules were mounted on in each session.

Loading

from datasets import load_dataset
ds = load_dataset("FeelAuthors/FEEL-Benchmark",
                  data_files="20260807/water_p25_collectorB_1_lerobot/data/chunk-000/*.parquet")

Or with LeRobot tooling, point a LeRobotDataset at a single trial directory.

License

CC BY 4.0.

Citation

@misc{feel_benchmark_2026,
  title  = {FEEL-Benchmark: Egocentric Tactile-Visual Manipulation Dataset},
  author = {FeelAuthors},
  year   = {2026},
  url    = {https://huggingface.co/datasets/FeelAuthors/FEEL-Benchmark}
}