Tidel AI · Bimanual Teleoperation Dataset

What We Sell›
PRODUCT

Real-machine teleoperation episodes. Each episode bundles synchronized multi-camera RGBD video, action commands, robot state, audio, calibration, and frame-level skill annotations.
UNIT OF SALE
- One episode= one complete sub-task demonstration
- Frame rate30 fps · hardware-synchronized streams
- AnnotationsBilingual (zh / en) shipped with every episode
SLICING DIMENSIONS
- By scene5 fuzzy buckets (see Catalog)
- By skill primitivepick · place · drop · pour · ... (from task_info.json)
- By arm usagesingle-arm / bimanual (per-span arm field)
- By contactfree-space vs wrench-bearing (state/end/wrench)
- By languageaction_text + english_action_text both shipped
DELIVERABLE FORMAT
- Per-episode treecamera/ + audio/ + parameters/ + proprio_stats/ + task_info.json
- Open formatsH.264 MP4 · FFV1 MKV · PCM WAV · HDF5 · JSON — no SDK lock-in
What's in the Catalog›
SCALE
- Total volume6,543 task hours
- Coverage9 level-1 scenes · 36 sub-tasks · 42 (scene, sub-task) pairs
- Hardware2 generations of the same wheeled bimanual platform
- Capture30 fps · multi-camera synchronized · bilingual annotations
DISTRIBUTION (BY HOURS)
- Home / Domestic2,916 h · 44.6%
- Office1,690 h · 25.8%
- Factory / Industrial830 h · 12.7%
- Healthcare572 h · 8.7%
- Retail / Service535 h · 8.2%
REPRESENTATIVE SUB-TASKS

Sort electronic products · Fold jeans · Sort items · Box / shelf placement · Cleanup & storage · Cable management · Pharmacy inspection · Garbage handling · Building blocks · Bedding · Pour water · Tableware · Item delivery · Floor sweeping · Books on desk · Remote control
What's Inside›
CAMERAS
- camera/video/camera_head_color.mp41280×720 · 30fps · H.264 · pre-rectified
- camera/video/camera_left_color.mp4640×360 · 30fps · H.264
- camera/video/camera_right_color.mp4640×360 · 30fps · H.264
- camera/depth/camera_head_depth.mkv1280×720 · 16-bit · FFV1 · aligned to color
- camera/depth/camera_left_depth.mkv640×360 · 16-bit · FFV1
- camera/depth/camera_right_depth.mkv640×360 · 16-bit · FFV1
AUDIO
- audio/audio.wav16kHz mono · PCM s16le
CALIBRATION
- parameters/{head,left,right}_intrinsic_params.jsonfx, fy, cx, cy, k1–k3, p1, p2
- parameters/{head,left,right}_extrinsic_params.jsonposition xyz · orientation quat
ACTION (TELEOP COMMANDS)
- action/joint/position[N, 14] · dual-arm joint targets
- action/effector/position[K, 2] · gripper command events
STATE (ROBOT OBSERVATION)
- state/joint/{position, velocity, effort}[N, 14] · arm joint readings
- state/effector/{position, velocity, force}[N, 2] · gripper readings
- state/end/{position, orientation, velocity, angular}[N, 2, 3-4] · EE 7-DoF pose
- state/end/wrench[N, 2, 6] · EE 6-axis force/torque
- state/head/{position, velocity, effort}[N, 2] · head pan/tilt
- state/waist/{position, velocity, effort}[N, 4] · waist 4-DoF
- state/robot/{position, orientation}[N, 3-4] · base SE(3) odometry
SYNC TIMESTAMPS · int64 ns
- timestampsmain timeline
- state_arm_joint_timestampsarm sample times
- state_effector_timestampsgripper sample times
- state_end_timestampsEE sample times
- state_head_timestampshead sample times
- state_waist_joint_timestampswaist sample times
- odom_timestampsbase odometry sample times
- hand_left_color_mp4_timestampsper-frame for left wrist video
- hand_right_color_mp4_timestampsper-frame for right wrist video
TASK METADATA
- task_info.jsonepisode_id, robot SN, init scene, action spans (skill, frames, bilingual text)
How to Evaluate›
SAMPLE PROVISION

A 1-hour sample snippet is available on access request (see sidebar). Sample ships in the same format as production delivery — same directory tree, same files, same annotation shape.
EVALUATION CHECKLIST · 4 dimensions
- 1. SYNC INTEGRITY9 timestamp arrays per episode, locked to a single master timestamps[] (int64 ns). Verify per-stream drift across the master clock.
- 2. ANNOTATION QUALITYtask_info.json → label_info.action_config[] carries: skill primitive, action_text + english_action_text, start_frame + end_frame, is_mistake flag.
- 3. ACTION ↔ STATEPlot proprio_stats.hdf5 → action/joint/position [N,14] vs state/joint/position [N,14] for any episode to inspect teleop tracking.
- 4. CALIBRATIONparameters/{head,left,right}_{intrinsic,extrinsic}_params.json. Reproject head depth → 3D point cloud or verify EE pose in camera frame.
What It's Like to Work With Us›
ENGAGEMENT MODEL
- Point of contactDedicated per account
- CommunicationShared channel (Slack / Lark) for active customers
- Delivery digestEpisode counts and per-episode QA flags surfaced with each batch
CUSTOM COLLECTION
- Bespoke scenesAccepted on quote
- Pre-captureJoint shotlist sign-off before recording begins
- Retake loopEdge cases re-captured against agreed criteria
OPERATIONS
- Capture methodReal-machine teleoperation by trained operators
- Quality flagsis_mistake surfaced in delivery manifest, not hidden
TYPICAL ENGAGEMENT FLOW

Sample evaluation → pilot bundle → rolling production delivery.
License Terms›
USAGE RIGHTS
- TrainingPerpetual license for model training and derivative weights
- Internal useResearch, evaluation, and product deployment permitted
- CommercialDeployment of trained models permitted
RESTRICTIONS
- RedistributionNo resale or redistribution of raw episode data
- OwnershipNo claim of ownership over the episodes themselves
- OperatorsNo reverse-identification of human operators
DATA SUBJECTS
- NDAAll operators are signed under non-disclosure agreements
- PrivacyLicense revocable on demonstrated operator privacy violations
EXCLUSIVITY
- StandardNon-exclusive — same episodes may be licensed to multiple buyers
- Exclusive packsAvailable on quote — commits provider not to re-license that subset
Delivery & Acceptance›
FILE STRUCTURE · per episode
- camera/video/H.264 MP4 — 1× head 720p · 2× wrist 360p
- camera/depth/FFV1 MKV — 16-bit gray, 3 streams
- audio/PCM WAV — 16 kHz mono
- parameters/6 JSON — 3× intrinsic + 3× extrinsic
- proprio_stats/Single HDF5 with action/* and state/*
- task_info.jsonEpisode metadata + action span labels
VOLUME
- Per episode (full)~300 MB at typical 30-second length (RGBD + proprio + audio)
- Per episode (color-only)~16 MB if depth streams omitted
- Catalog scaleTB-class — transfer mechanism scoped per contract
DELIVERY
- ChannelDefined per contract
- ScheduleBulk on first delivery; rolling for subscriptions
- ManifestPer-episode QA report shipped with each batch
ACCEPTANCE
- Pass criteriaSync OK · annotation present · is_mistake = false
- Failed episodesRe-collection or hour-credit at customer choice

Secret access key required, provided by us upon request.

FACTORY

LOGISTICS

WAREHOUSING

SERVICES

WORKPLACE

HEAD · COMBINED

LEFT WRIST

RIGHT WRIST

EPISODESone per object

SUBTASKSpick · place

PHASESfrom gripper signal

PRIMITIVESreach·grasp·transport·place

KEYFRAMES[ ] to step

PRIMITIVES14-primitive taxonomy

RIGHT ARM 7 joints · rad state action

RIGHT GRIPPER
pos · force

BASE POSE x · y · yaw yaw x,y

ARM SPEED
Σ|q̇| · rad/s

SHIPPED WITH EPISODE RGB × 3 (1× head RGBD · 2× wrist RGB) DEPTH × 3 (16-bit FFV1) AUDIO 16kHz mono ACTION joint × 14 ACTION gripper × 2 STATE joint × 14 EE pose 7D × 2 EE wrench 6D × 2 GRIPPER force × 2 BASE odometry SE(3) WAIST 4-DoF HEAD 2-DoF 9× hardware-sync timestamps INTRINSIC + EXTRINSIC × 3

00:00 / 00:28

EPISODE

ROBOT

SCENE

INIT

HOW THESE LABELS WERE MADE every layer is traceable to a signal or a frame

26 KEYFRAMES · HEAD CAMERA OVERLAYS red = component A · white = component B · yellow/green = bins