Sample data outputs · Stereo + wrist cams
Egocentric Stereo + Wrist Cam Example Videos
Last reviewed:
Short answer
Egocentric stereo plus wrist-cam example videos show one task through every stream at once: a head-mounted stereo pair, a camera on each wrist, depth, three IMU traces, hand-pose overlays and an action-segment timeline. Four real clips cover refrigerator assembly, control panel and PCB install, material handling and commercial cleaning.
- Views
- Head stereo pair + two wrist cameras
- Motion
- 200 Hz 6-DoF IMU on head and both wrists
- Sync
- One PTP clock domain
- Hands
- 21 joints per hand, 3D
Example videos
What do egocentric stereo + wrist cam recordings look like?
Four of our own recordings, each shown as a synced multi-stream visualization. The two hardest, most hand-intensive tasks come first. Use the jump links to open the player at the highlighted segments.
Clip 1 · Refrigerator assembly · 4:16
Tapping and adjusting refrigerator components
The worker uses a mallet to tap the refrigerator frame and seat the door, making final adjustments to components.
What to watch for
- The tap-and-seat sequence, where the door is seated to the frame by tapping and adjusting.
- Whether the head view keeps the contact point in sight while the hands are close, then what the wrist views show at the same moment.
- How the action-segment timeline moves through the task as the work changes.
Clip 2 · Control panel and PCB install · 4:13
Securing internal wires with ties
A control panel and PCB install, with connector mating and wire securing worked by hand at close range.
What to watch for
- Connector mating and wire securing in the two highlighted segments.
- Small parts held close to the lens: compare the head stereo pair with the left and right wrist views.
- The hand-pose overlays on both hands during close work.
Clip 3 · Material handling · 4:56
Pushing carts with materials
Warehouse and material-handling footage of carts being pushed with materials.
What to watch for
- Material-handling work with carts and loads, shown from the head stereo pair and both wrist cameras.
- The three IMU traces (head, left wrist, right wrist) across the whole clip.
- The depth panel across the 0.25–2.5 m colour scale.
Clip 4 · Commercial cleaning · 12:00
Cleaning with a mop
Commercial cleaning crew work, including window work, in a single 12-minute recording.
What to watch for
- The longest of the four clips, with mopping and window work in one recording.
- How the action-segment timeline divides a long, repetitive task into segments.
- The wrist views during mopping and window work.
The ODOM 3D panel that appears in these visualizations is shown as a panel only. Head or camera pose, odometry and SLAM are not stated as deliverables on this site.
Panels
What does each panel in the videos show, and what is delivered?
Every video is a synced visualization of the same episode streams. The table maps each on-screen panel to the stream behind it and to the field the site states for it. Anything the site does not state is marked Not stated.
| On-screen panel | Stream | Delivered field, as the site states it | Not stated |
|---|---|---|---|
| EGO LEFT + EGO RIGHT | Head-mounted stereo pair at a human interpupillary baseline, calibrated per session | head_rgb; stereo pairs delivered raw with calibration, rectified pairs and disparity on request | How left and right are split inside the head_rgb field |
| WRIST LEFT + WRIST RIGHT | One camera on each wrist, 1080p/60 | wrist_l and wrist_r, delivered as their own streams alongside the head view | Nothing further |
| DEPTH (0.25–2.5 m colour scale) | Depth, 848×480 at 30 fps | depth_mm (metric millimetres) and depth_conf per-frame confidence | Which sensor produced the depth in these four clips; the colour scale is a display choice |
| IMU HEAD, IMU WRIST L, IMU WRIST R | 6-DoF IMU at 200 Hz on the head and both wrists, accel + gyro | imu_head, imu_wrist_l and imu_wrist_r | Nothing further |
| Per-frame timestamp and frame counter | All streams hardware-triggered onto one PTP clock domain | Per-stream nanosecond timestamps (*_timestamp_ns) | A frame counter as a delivered field |
| Action-segment timeline | Action segments with plain-language labels shown on the timeline | action_segments, stored as start_ns, end_ns, verb, noun | How the on-screen label text maps onto verb and noun |
| Hand-pose overlays | 21 joints per hand, 3D, in metres in the head-camera frame | hand_joints and a per-joint hand_visible flag for occluded fingers | Nothing further |
| ODOM 3D trajectory panel | A panel visible in the visualization | Not stated. The site does not state head or camera 6-DoF pose, odometry or SLAM as a deliverable | Whether it is delivered, and in what format |
| Delivery formats | RLDS, LeRobot, WebDataset, HDF5 or zarr, all projections of one episode schema. See the episode schema and delivery formats. | ||
Why stereo + wrist cams
Why use stereo and wrist cameras for hard, skilled tasks?
Skilled assembly puts hands, tools and small parts exactly where a single head camera cannot see them. Each stream below answers a specific gap.
Close hand work hides the view from the head
In skilled assembly the hands, a tool or a part often sit between the head camera and the thing being worked on. The head camera still loses sight of a grasp when the hand blocks it, which is why wrist views and hand pose travel with the stereo pair.
Wrist views sit next to the contact
A camera on each wrist looks at what the hand is doing: a connector being mated, a tie being pulled tight, a fastener being seated. The clips above let you check that for yourself, panel by panel.
Stereo depth is measured, not guessed
Depth from a single frame is a guess, and a wrong guess about where a part sits costs a failed insertion. A calibrated stereo pair measures disparity between two views taken at a fixed, known baseline, with intrinsics, extrinsics and the baseline measured every session and shipped with the episode.
One clock keeps every panel in step
Stereo, depth, IMU and wrist views are hardware-triggered onto one PTP clock domain, so there is no drift to correct between sensors. The per-frame timestamp in each video is that shared timeline made visible.
FAQ
Questions about the stereo + wrist cam examples.
- What do the four example videos show?
- Four recordings of skilled and physical work, each shown as a synced multi-stream visualization: refrigerator assembly (4:16), control panel and PCB install (4:13), material handling with carts (4:56) and commercial cleaning with a mop, including window work (12:00). The refrigerator and PCB clips come first because they are the hardest, most hand-intensive work.
- What is in each synced panel view?
- Each video shows an ego left and ego right head-mounted stereo pair, a wrist left and wrist right camera, a depth panel on a 0.25–2.5 m colour scale, three IMU traces (head and both wrists, accel and gyro), a per-frame timestamp and frame counter, an action-segment timeline with plain-language labels, and hand-pose overlays. The panel-to-field table on this page maps each one to the field the site states.
- Why add wrist cameras to a head-mounted stereo pair?
- The head camera still loses sight of a grasp when the hand blocks it. A camera on each wrist keeps the contact, connector or fastener in view from the hand itself, and hand pose travels with the pair. The stereo pair adds measured depth from a calibrated, fixed baseline.
- Do the videos show head or camera pose, odometry or SLAM?
- The visualizations include an ODOM 3D trajectory panel, and this page describes it only as a panel visible in the video. The site does not state head or camera 6-DoF pose, odometry or SLAM as a deliverable, so this page does not promise it. The stated motion stream is the 200 Hz 6-DoF IMU on the head and both wrists. If you need more, put it in the skill spec; what is delivered for your order is agreed there.
- How is the data delivered?
- Episodes ship as RLDS, LeRobot, WebDataset, HDF5 or zarr, all projections of one canonical episode schema with meta, streams, annotations and calibration groups. Every stream carries nanosecond timestamps in the PTP clock domain, so sync can be verified independently.
- How do I request a custom stereo and wrist-cam dataset or pilot?
- Describe the task, its steps, how each step starts and ends, and the failure cases your model is missing. That becomes a written skill spec, quoted before any capture. Most programs begin with a Spec Pilot, a small, fast first batch that proves the spec on real footage before volume scales. Request a custom dataset and a rep will scope it with you.
Want this setup on your task?
Send the task, its steps and the failure cases your model is missing. We quote a written skill spec before any capture, usually starting with a Spec Pilot.