Data return schemeSolutionsEnglishE2 Product Page
E2 · EPISODE 20260917_133640
Device UID d36997d92fc9, serial number 126091102913. The skeleton, mosaic, and color block in the upper right corner of the player are model estimates, not the hand, pose, or depth output of this device. What the device actually contains remains the left and right cameras, IMU, magnetometer, and audio. This package does not include head pose and SLAM; the heading at the bottom of the frame is gyro integration and will drift.
E2 · EPISODE 20260917_133640
Device UID d36997d92fc9, serial 126091102913. The skeleton, mosaic and corner patch are model estimates, not device hands, pose or depth. The device files are still the two eyes, IMU, magnetometer and audio. This package has no head pose and no SLAM. The heading in the bar is integrated gyro and it drifts.
Buffering this segment
Both channels are 16 kHz stereo audio from the device, starting 97 ms later than the left camera's exposure start. When played together, only the channel you press sounds, avoiding overlap. On mobile, the annotated clip appears first. The mosaic covers faces, distant human figures, text with confidence reaching 0.40, bright screens, and the blue wall marker.
Source 0:00 · 0.00 s · Frame 0 · 30 Hz
No record has been loaded for this second yet.
No record for this second yet.
The blue sign on the wall is masked by color-connected region: at least 160 pixels wide, blue pixel ratio approximately 0.15 to 0.55, located in the upper half of the frame. Around 10:28 there is one block in each eye view. Such blocks appear 32 times across the entire recording, not just in this one second. This is not manually painted on a single frame.
OCR is split into two tiers. Only results reaching 0.40 are adopted. The full segment repeatedly read correctly as SKYWORTH. Around 2:17, the cropped logo also read "Skyworth XR". At 10:28, the color redaction is still in place, but for that second the OCR fell back to a low-confidence misspelling, e.g. SKYWORH. Most other results reaching 0.40 are fragmented characters and are not treated as object names. Faint text on the glass was not separately redacted.
Hands follow the Vision skeleton in the player, not the device joints. Faces and distant human bodies are only masked when detected. Cartons, doors, shoes, and ceilings have no category boxes. The gyroscope timeline only indicates head-mounted angular velocity, not action names.
The blue wall mark is masked by a color region: at least 160 pixels wide, blue-pixel share about 0.15 to 0.55, in the upper half of the frame. At 10:28 both eyes have one region. The log has 32 such rows, not only that second. It is not a hand-painted patch.
OCR is accepted only at confidence 0.40 or above. The string that repeats is SKYWORTH. Around 2:17 the crop of the mark also reads Chuangwei XR. At 10:28 the color mask is still on, but OCR falls back to low-confidence errors such as SKYWORH. Other accepted strings are fragments, not object names. Faint glass lettering is not separately masked.
Hands in the player are a Vision skeleton, not device joints. Faces and distant people are masked only when detected. Boxes, the door, shoes and the ceiling have no class boxes. The gyro timeline is head-worn angular rate, not an action name.
The timeline is still gyroscope magnitude, not action names. The "this second" above comes from model records: hands, blue wall markers, adopted text, low-confidence OCR, faces, bodies, and bright screens. Cardboard boxes and doors have no class boxes. The faint text on the glass was not separately annotated this time.
The timeline is still gyroscope magnitude, not an action name. The line above comes from the model log: hands, blue wall marks, accepted text, low-confidence OCR, faces, people and bright screens. Boxes and the door have no class labels. Faint glass lettering is not separately masked.
time_sync.jsonstereo_index.jsonlmotion_spans.jsonlnarration.jsonle2-findings.jsonl
STEREO
Camera0 is left, Camera1 is right. HEVC 1600×1200. The calibration group is labeled tracking. The p50 of the left-right exposure start offset is 198 microseconds, and the p50 of the exposure midpoint offset is 210 microseconds. Exposure duration is approximately 4–10 ms, with gain fixed at 1024. Pinhole estimate of horizontal FOV is approximately 117°, without distortion coefficients — not a product specification.
IMU + MAG
Accelerometer and gyroscope at approximately 798.9 Hz; mean accelerometer magnitude spot-check of 9.83 m/s². Magnetometer at exactly 200 Hz. The accelerometer starts 124 ms before the exposure start of frame 0 of the left camera. The 1.5 ms camera alignment in the calibration was not applied because this package has no head pose to verify against.
AUDIO
PCM 16 kHz stereo. Packet 0 is 96.6 ms later than the left-eye exposure start. Do not hard-align audio.mp4 and video from 0 seconds.
NOT IN THIS PACKAGE
The device files contain no hand joints, head pose, SLAM, depth, or object categories. The hands in the player are Vision estimates, the upper right is disparity rather than a depth camera, and the bottom heading is not SLAM.
A · DEVICE · CAN TRAIN
Training reads the raw 30 Hz footage from Camera0 and Camera1, totaling 20,466 frames, with 0 gaps exceeding 40 ms. Do not train with this 4 Hz preview. This is a human first-person observation,not_robot_action is true, and cannot be treated as a robot end-effector action.
Train on the 30 Hz Camera0 and Camera1 files, 20466 frames, with no gap longer than 40 ms. Do not train on this 4 Hz preview. This is a human egocentric observation. not_robot_action is true. It is not a robot end-effector action.
The 65.94 mm baseline comes from the extrinsic position difference, not a checkerboard. The pinhole horizontal field of view is approximately 117°, with no distortion applied. Until calibration is completed, only relative stereo is performed — no metric depth.
The 65.94 mm baseline is an extrinsics position difference, not a checkerboard. Pinhole horizontal field is about 117°. Distortion was not used. Until a new calibration, use relative stereo only, not metric depth.
The IMU integration start is 124.3 ms earlier than the left-eye exposure start. Gyroscope RMS median over a 5-second window is 0.352 rad/s. Magnetometer at 200 Hz; indoor magnetic interference was not separately tested. The factory alignment constants are in time_sync.json and were not applied on this page, because there is no head pose available for verification.
Accel leads the left-eye exposure start by 124.3 ms. Median 5 s gyro RMS is 0.352 rad/s. Magnetometer is 200 Hz. Indoor magnetic disturbance was not measured. Factory alignment constants are in time_sync.json and were not applied here: there is no head pose to check them.
Stereo packet 0 is 96.6 ms late. Weak alignment such as keystrokes and speech is possible, but audio and video cannot be hard-aligned from 0 s.
Stereo packet 0 is 96.6 ms late. It can weakly align key clicks or speech. Do not stack audio and video at t=0.
Hard synchronization between left and right eyes, exposure start offset p50 198 µs, suitable for representation pre-training. Hand, screen, and person bounding boxes are not labels at this layer.
The eyes are hard-synced. Exposure-start difference is 198 µs at the median. That supports representation pretraining. Hands, screens and people are not labels of this layer.
B · MODEL ESTIMATE · NOT GROUND TRUTH
Vision records from annotated clips, one row per second. The hand is a 21-point estimate, not device joints. The bright screen is a brightness grid, not a display detector. These numbers cannot be used as ground truth.
These rows come from the annotated preview, one per second. Hands are a 21-point estimate, not device joints. Bright screens are a luminance grid, not a monitor detector. Do not treat the counts as ground truth.
Blue wall label 32 seconds, 57 frames in total. The text "SKYWORTH" appears 8 times. In the recorded weak fields, "Skyworth XR" reaches a confidence of 0.40 12 times, and the on-page "this second" is treated as adopted. Most others reaching 0.40 are fragmented characters, not object names. Cardboard boxes, doors, and keyboards/mice have no category boxes.
Blue wall marks: 32 seconds, 57 regions. SKYWORTH appears 8 times in the accepted-text field. In the weak field, Chuangwei XR reaches confidence 0.40 twelve times, and the per-second line treats that as accepted. Other strings at 0.40 are fragments, not object names. Boxes, the door, the keyboard and the mouse have no class boxes.
The following four segments were manually reviewed in the annotation preview, or gyroscope-only. Tap to jump to that second. These are not detectors, nor full-length action names.
The four windows below are a human look at the annotated preview, or gyro only. A click seeks there. They are not a detector and not action names for the whole take.
C · STILL MISSING FOR A ROBOT POLICY
The papers below are other people's requirements for data fields, not the experimental results of this segment, nor the accuracy measured on this device.
The papers below are other groups' field requirements. They are not results of this episode and not a measured accuracy of this device.
Open X-Embodiment / RT-X (arXiv 2310.08864) binds observations together with robot actions. This section contains only human observations, with no end-effector actions, contact forces, or language instructions.
Open X-Embodiment / RT-X (arXiv 2310.08864) pairs observations with robot actions. This take has a human observation only. No end-effector action, contact force, or language instruction.
HOT3D (CVPR 2025, arXiv 2406.09598) requires 3D hands and objects. This Vision 21-point segment has no per-joint confidence, nor any glove or MANO.
HOT3D (CVPR 2025, arXiv 2406.09598) requires 3D hands and objects. The Vision 21 points here have no per-joint confidence, and there is no glove or MANO.
Ego-Exo4D (CVPR 2024) uses dense narration. This segment of narration.jsonl is sampled approximately every 60 seconds; the method states manual sampling, not model output.
Ego-Exo4D (CVPR 2024) uses dense narration. The narration.jsonl here is a human sample about every 60 seconds, not a model output.
Navigation and multi-step manipulation require head pose or VIO. This segment does not have it. The bottom heading is debiased gyroscope integration; the axes have not been verified against the camera and will drift.
Navigation and multi-step manipulation need head pose or VIO. This take has neither. The heading is bias-subtracted integrated gyro. Axes were not checked against the camera, and it drifts.
The color block in the upper right is coarse SAD on an uncorrected fisheye image, not converted to meters using focal length and baseline. Grasp pose requires a separate checkerboard or depth camera.
The corner patch is coarse SAD on unrectified fisheye. Focal length and baseline were not converted to meters. A grasp pose requires a new checkerboard or a depth camera.
The annotation clips cover detected faces, distant human bodies, text above 0.40, bright screens and blue wall marks. 77 seconds of faces, 221 seconds of human bodies and 642 seconds of bright screens indicate that these are common in office footage. Faint text on glass may still remain. This is not a clean stream ready for direct trading.
The annotated preview masks detected faces, distant people, text at 0.40 or above, bright screens and the blue wall mark. 77 face seconds, 221 person seconds and 642 bright-screen seconds show how often they appear. Faint glass lettering can remain. This is not a clean file ready to sell as fully masked.
FOUR USERS · ONE EPISODE
The same segment 20260917_133640. The only difference is what you want to train with it. The numbers on the sample card all come from this segment's files, not from product specifications.
The same episode, 20260917_133640. The difference is what you train. Numbers on the cards come from this episode's files, not from a brochure.
For those working on self-supervised representation learning, binocular vision, and temporal alignment. This segment can go into the loader right now. It is a human observation, not a robot action.
For people training self-supervised video, stereo, and time alignment. This take can enter a loader now. It is a human observation, not a robot action.
For those who will build skill timelines later. Below are candidate windows, not already-labeled action names, nor frame-by-frame detections.
For people who will build a skill timeline later. The windows below are candidates, not finished action names and not a per-frame detector.
For those doing stereo matching. This segment can deliver relative geometry. Metric depth is not yet available because the baseline was not calibrated with a checkerboard.
For people running stereo matching. This take can hand over relative geometry. Metric depth is not here, because the baseline was not measured on a checkerboard.
For those who need to confirm whether an order can be placed and whether the fields are clearly written. This page is the data card. The mosaic on the annotated clip is not a fully redacted, clean stream.
For people checking whether the fields are clear enough to order. This page is the data card. The mosaic on the annotated preview is not a fully cleaned stream.
All four user types follow the same rule: read-only access to Camera0 / Camera1; IMU is 124.3 ms early, audio is 96.6 ms late; skeleton and OCR are labeled as model estimates; do not treat this segment as robot action supervision.
All four users share one rule: read only Camera0 and Camera1; IMU is 124.3 ms early and audio is 96.6 ms late; mark skeletons and OCR as model estimates; do not supervise a robot action with this take.