Short answer
Egocentric video, recorded from a head- or chest-mounted camera, shows hands and objects up close from roughly where a humanoid's head camera sits, which suits hand-object learning, hand tracking and co-training. Exocentric video, from fixed or tripod cameras, shows posture, workspace layout and the whole task, which suits step recognition, timing and whole-body context. Many programs need both, recorded in sync.
The viewpoint decision behind every video purchase
Footage a business already holds almost always comes from a single viewpoint. Technician body cameras and smart-glasses sessions are egocentric; training films shot by a camera operator, process-monitoring cameras and CCTV are exocentric. When you commission new recording you choose, and that choice fixes which models the footage can feed for the life of the dataset.
The useful question is not which view is better, but which parts of your stack each one serves and how far each sits from the cameras your humanoid will carry. For the general limits of human footage in robot learning, see what human demonstration video can and cannot do.
What each viewpoint captures
| Property | Egocentric (head or chest camera) | Exocentric (fixed, tripod or ceiling camera) |
|---|---|---|
| Hands and grasped objects | Large in frame; fingers often resolved at bench distance | Small, often hidden by the worker's torso or arms |
| Whole-body posture and stepping | Mostly out of frame; inferred only from head motion | Visible, including reach, crouch and carry posture |
| Workspace layout and other people | Partial; changes with every head turn | Stable, wide context of stations and traffic |
| Camera motion | Continuous head motion; blur on fast turns | None, or deliberate pans; stable background |
| Match to a humanoid head camera | Similar direction, but human head motion is larger and faster | Poor, unless the deployment cell also has fixed cameras |
| Common existing sources | Body cameras, remote-assistance glasses, some point-of-view training clips | Training films, process cameras, CCTV |
Two consequences matter for procurement. Egocentric footage loses the hands whenever the wearer looks away, so hand visibility varies widely by task and must be measured, not assumed. Exocentric footage keeps the whole task in view but rarely resolves finger-level detail, so it answers "which step is happening" far better than "how is the part being held".
What published work uses each view for
Most robot-learning research on human video builds on the first-person view. Ego4D collected 3,670 hours of daily-life video from 931 camera wearers in 74 locations across 9 countries, with consenting participants and de-identification procedures (Grauman et al., 2021). EPIC-KITCHENS-100 adds 100 hours from head-mounted cameras in 45 kitchens with dense action labels (Damen et al., 2020). Co-training work also relies on this view: EgoMimic recorded demonstrations with Project Aria glasses and trained one policy on human and robot data (Kareer et al., 2024), and Humanoid Policy ~ Human Policy used hand and finger poses captured with consumer VR devices on tasks aligned with the robot's (Qiu et al., 2025).
Exocentric views earn their place where the whole scene matters. Ego-Exo4D recorded skilled activities simultaneously from an egocentric camera and several exocentric cameras: 740 participants, 13 cities, 123 scene contexts and 1,286 hours in its arXiv v4 release, with benchmarks that include fine-grained activity understanding and proficiency estimation (Grauman et al., 2023). Assembly101 used 8 static and 4 egocentric cameras to record assembly and disassembly (Sener et al., 2022). The views also complement each other: affordance grounding has been transferred from exocentric images of people using objects to egocentric images of the objects alone (Luo et al., 2022).
Robot datasets mix viewpoints too. DROID records every episode with two exterior stereo cameras plus a wrist-mounted camera (Khazatsky et al., 2024). If your policy consumes a head camera and wrist cameras, human footage that approximates those views is easier to use than footage from the ceiling.
Which view serves which part of the stack
| Use in your stack | Lean toward | Reason |
|---|---|---|
| Visual pretraining for manipulation | Egocentric | Hands and objects dominate the frame, as in robot head and wrist views |
| 3D hand tracking, co-training, retargeting | Egocentric, ideally with headset hand tracking | Fingers are resolved; headset devices can log hand poses directly |
| Step recognition and task progress | Exocentric, or both | Stable framing makes boundaries easier to see and label |
| Posture, carrying, walking with loads | Exocentric from several angles | Legs, torso and reach stay visible |
| Proficiency and technique judgments | Both, synchronized | Experts judge posture and hand detail together |
| Evaluating task-understanding models | Whatever matches deployment cameras | An evaluation set should look like what the model will see |
Matching the view to your humanoid's cameras
A head-mounted camera is a starting point, not a match. A human wearer's eyes sit at a different height from most humanoid head cameras, the head turns faster and further than a robot neck, and consumer lenses are often wider or narrower than the robot's. Record camera model, lens, mount position and orientation per session so you can crop, rectify or reweight later; the method is set out in matching demonstration camera viewpoints to a humanoid's sensors, and the recording parameters in the head-mounted camera capture specification.
For exocentric footage the deciding question is whether the deployment site will have fixed cameras at all. If the robot will be the only observer, exocentric footage helps with understanding and labeling but contributes little to the policy's own observations.
Existing footage or commissioned capture for each view
Existing egocentric footage comes mainly from field-service body cameras and remote-assistance sessions. These often carry expert narration but have inconsistent framing and incomplete task coverage. Existing exocentric footage is plentiful but seldom resolves manipulation; fixed-camera and CCTV footage usually supports only activity, timing and exception detection.
Synchronized ego-exo recording almost always means commissioned capture, because every camera needs a known position, calibration and a shared clock. The rig is covered in multi-view synchronized recording of workplace tasks, and what the delivery must include is set out in calibration requirements. For the wider buy-or-commission trade-off, see existing workplace footage vs commissioned recording.
Illustrative capture plan for paired ego-exo sessions
# Illustrative only — example values, not an available dataset or a recommendation
session: station_04_kitting # synthetic ID
egocentric:
device: head-mounted RGB camera with IMU # model and serial recorded per session
frame_rate: constant 30 fps
hands_visible_target: buyer sets per task; measured on delivery
exocentric:
cameras: 3 fixed (front-high, side, overhead)
frame_rate: constant 30 fps
placement: extrinsics logged relative to a station origin marker
sync:
method: shared clock plus visual and audio slate at start and end
tolerance: buyer sets (for example, within one frame)
calibration: intrinsics per camera; extrinsics per session; head-camera pose track
labels: task and step segments on an exocentric master; hand boxes on egocentric
not_included: robot actions, joint states, contact forces
Limitations of both views
- Neither view contains robot actions, joint states or forces; contact-rich tasks need extra data, as explained in when video is not enough.
- Egocentric video suffers from self-occlusion by the wearer's arms, motion blur during head turns and framing that follows attention rather than the task.
- Exocentric video loses finger detail and depth unless cameras are close, high-resolution or depth-capable.
- Extra views multiply consent scope, storage and annotation effort, and they are only useful together if synchronized and calibrated.
- Gloves, PPE and production pace degrade both views.
Questions to ask suppliers
- Which camera types and mounts appear in the footage, and is that recorded per clip?
- For egocentric clips, what share of task time shows both hands, measured on a random sample?
- For exocentric clips, how far is the camera from the work surface, and how many pixels span a hand?
- Were any sessions recorded from more than one view at once, and how were the views synchronized?
- Are intrinsics and camera positions available, or can they be recovered at the site?
- Who appears in the exocentric frame besides the demonstrator, and what notice did they receive?
- Do the egocentric devices log head pose, IMU, gaze or hand tracking, and are those streams licensable?
Discuss your requirements
If you are choosing between first-person footage, fixed-camera footage or a paired rig, describe the views, tasks and permitted uses to SourceX. SourceX reviews the requirement, investigates whether US businesses hold matching material, checks the supplier's licensing permissions and follows up about potential sources; if no existing asset fits, SourceX can scope a commissioned collection of new recordings. A request does not guarantee a match. The humanoid training data guide and the workplace task demonstration hub cover the other routes, and the catalog entry for first-person video of skilled manual work describes a related asset type.
Discuss your humanoid training-data requirements with SourceX
Sources
- Ego4D: Around the World in 3,000 Hours of Egocentric Video — scale and composition of a large egocentric corpus with consent and de-identification.
- Rescaling Egocentric Vision — EPIC-KITCHENS-100 head-mounted kitchen recordings with dense action labels.
- EgoMimic: Scaling Imitation Learning via Egocentric Video — egocentric demonstrations from Project Aria glasses co-trained with robot data.
- Humanoid Policy ~ Human Policy — egocentric hand and finger poses from consumer VR devices as humanoid training data.
- Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives — simultaneous egocentric and exocentric recording of skilled activity.
- Assembly101: A Large-Scale Multi-View Video Dataset for Understanding Procedural Activities — static plus egocentric camera rig for assembly tasks.
- Learning Affordance Grounding from Exocentric Images — transfer of affordance knowledge from exocentric to egocentric views.
- DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset — robot episodes recorded with exterior and wrist cameras.