SoTac — SO-101 Tactile Manipulation Corpus

SoTac is a growing corpus of visuo-tactile teleoperation demonstrations recorded on an SO-101 arm with dual 3-axis tactile fingertip sensors. Every 30 Hz observation carries a full per-taxel force map for both fingertips, and every episode ships with the raw ~91 Hz tactile stream as a sidecar — no downsampling at record time.

Created using LeRobot (via the lerobotac sensor fork).

Use the button above — it opens the LeRobotAC tactile visualizer with per-taxel 3D force arrows, contact timelines, and the raw 91 Hz stream panels. The standard LeRobot viewer does not render the tactile data.

Contents

SoTac is actively growing — new tasks, objects, and sensors will be added over time. This is the curated set: every episode was human-reviewed. Episodes with contact faults (accidental gripper contact during approach, closed-finger approaches, post-release contact), operator visibility, or sensor faults are removed, and idle heads/tails are trimmed so every episode starts near task onset. The unedited archive is Jingyi-Z/sotac_raw.

Current release (2026-08-26): 63 episodes / 25,401 frames (~14 min) across 3 pick-and-place tasks, grouped by task:

task	episodes	episode #
Pick up the red foam ball and place it into the container	21	0–20
Pick up the orange rubber ball and place it into the container	22	21–42
Pick up the red cup and place it on top of the container	20	43–62

Per-episode annotations ship in annotations/:

• episode_annotations.json: episode-level curation labels — grasp type, attempt count, outcome (success / failure / partial), distractor objects (11 episodes include an orange ball or white box as distraction), grip-force quality labels (insufficient/excessive force), and free-form event notes.
• episode_XXXXXX.json (all 63 episodes): time-aligned language annotations — subtask segments (approach, grasp, transport, place_release) and tactile event interjections (contact, grasp_stable, slip, rotation, place, release/drop) with per-event confidence. Auto-generated from the 91 Hz tactile stream and human-reviewed in the visualizer.

Stable provenance: every episode carries a source_raw_episode field — both as a column in meta/episodes and in the annotations — pointing to its index in sotac_raw, which is append-only and never renumbered. Curated releases may renumber freely; raw indices are the permanent episode identity. curation_map.json in the root holds the full mapping.

Grip forces span gentle (4–10 N peak) to firm (50 N peak) handling across sessions.

Data collection: episodes 0–5 & 43–47 by Jingming Zhang, episodes 21–33 by Yuzhou Wang, episodes 6–20, 34–42 & 48–62 by Jingyi Zou.

Curation provenance: derived from sotac_raw with curate_tactile_dataset.py and the spec curation_sotac_v2.json (same repo) — all deletions and per-episode keep-windows are reproducible.

Hardware

• Robot: SO-101 follower arm, teleoperated with an SO-101 leader arm
• Cameras: wrist + top, 640×480 @ 30 fps (MJPG)
• Tactile: 2× Paxini PX-6AX GEN3 (DP-S2015-Elite) fingertip pads on a shared High-Speed Communication Board — 52 taxels per pad, 3-axis force per taxel, 0.1 N/LSB, ~91 Hz effective rate for both pads together. Firmware zeroing at connect. Finger [0] = gripper-side pad (module 10), finger [1] = wrist-roll-side pad (module 18).

Data format

Standard LeRobot v3.0 dataset, plus one extra feature and one extra directory:

observation.sensors.paxini_fingertip — float32, shape (2, 52, 3): the latest raw tactile sample per finger at each 30 Hz observation (fingers × taxels × [fx, fy, fz], in units of 0.1 N). Per-taxel (x, y, z) mm coordinates for the GEN3 layout are available in paxini-sdk.

sensors/paxini_fingertip/episode_XXXXXX/ — raw ~91 Hz stream sidecars:

• sensor_1.csv, sensor_2.csv (fingers 0 and 1): 163 columns — timestamp_ns, frame_status, time_calibration_offset_ns, calibrated_timestamp_ns, fx, fy, fz, p_00_fx … p_51_fz (resultant force followed by 52 per-taxel triplets)
• alignment.json: episode-start epoch timestamp (time.time_ns) anchoring the raw stream to frame 0 of the 30 Hz main table

Collection protocol

Episodes start with the gripper open and the pads contact-free (calibration reference). The operator and leader arm are kept out of both camera views where possible; remaining appearances are flagged per-episode in the annotations. Object placement is varied between episodes; some episodes include distractor objects.

Tooling

• Recording: lerobotac (hall-sensor branch) — LeRobot fork with a sensor framework; the paxini sensor's combined output mode + record_raw_csv produced this dataset
• Sensor SDK: paxini-sdk
• Visualization / annotation: LeRobotAC Dataset Visualizer

Dataset Structure

meta/info.json:

{
    "codebase_version": "v3.0",
    "fps": 30,
    "features": {
        "action": {
            "dtype": "float32",
            "names": [
                "shoulder_pan.pos",
                "shoulder_lift.pos",
                "elbow_flex.pos",
                "wrist_flex.pos",
                "wrist_roll.pos",
                "gripper.pos"
            ],
            "shape": [6]
        },
        "observation.state": {
            "dtype": "float32",
            "names": [
                "shoulder_pan.pos",
                "shoulder_lift.pos",
                "elbow_flex.pos",
                "wrist_flex.pos",
                "wrist_roll.pos",
                "gripper.pos"
            ],
            "shape": [6]
        },
        "observation.images.wrist": {
            "dtype": "video",
            "shape": [480, 640, 3],
            "names": ["height", "width", "channels"],
            "info": {
                "video.height": 480,
                "video.width": 640,
                "video.codec": "av1",
                "video.pix_fmt": "yuv420p",
                "video.is_depth_map": false,
                "video.fps": 30,
                "video.channels": 3,
                "has_audio": false
            }
        },
        "observation.images.top": {
            "dtype": "video",
            "shape": [480, 640, 3],
            "names": ["height", "width", "channels"],
            "info": {
                "video.height": 480,
                "video.width": 640,
                "video.codec": "av1",
                "video.pix_fmt": "yuv420p",
                "video.is_depth_map": false,
                "video.fps": 30,
                "video.channels": 3,
                "has_audio": false
            }
        },
        "observation.sensors.paxini_fingertip": {
            "dtype": "float32",
            "shape": [2, 52, 3],
            "names": ["finger", "taxel", "axis"]
        },
        "timestamp": {"dtype": "float32", "shape": [1], "names": null},
        "frame_index": {"dtype": "int64", "shape": [1], "names": null},
        "episode_index": {"dtype": "int64", "shape": [1], "names": null},
        "index": {"dtype": "int64", "shape": [1], "names": null},
        "task_index": {"dtype": "int64", "shape": [1], "names": null}
    },
    "total_episodes": 63,
    "total_frames": 25401,
    "total_tasks": 3,
    "chunks_size": 1000,
    "data_files_size_in_mb": 100,
    "video_files_size_in_mb": 200,
    "data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
    "video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4",
    "robot_type": "so_sensor_follower",
    "splits": {
        "train": "0:63"
    }
}

Citation

@misc{zou2026sotac,
  author       = {Zou, Jingyi},
  title        = {SoTac: An SO-101 Tactile Manipulation Corpus},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Jingyi-Z/sotac}}
}