Short answer
Gloves and PPE change hand appearance, hide skin and finger contours, add bulk, and can degrade 3D hand-pose estimators trained mostly on bare hands. Workplace footage often shows gloves for good reasons, so buyers should measure tracking quality on gloved samples, decide whether gloved hands are the deployment target, and plan annotation or instrumented capture where estimates fail.
Why this matters in real workplaces
Most public hand-tracking data shows bare hands on tabletops. Real work does not. Food handlers wear nitrile gloves, warehouse staff wear cut-resistant gloves, laboratory staff wear double gloves, electricians wear insulating gloves, and construction crews wear heavy leather gloves with sleeves, safety glasses, hard hats and high-visibility clothing. Footage from operating businesses is full of these.
For a buyer, three questions follow. Can hands in this footage be tracked well enough to use? Does the gloved hand represent what the deployed robot will do, or does it add a nuisance? And what does the robot's own hand look like, since a humanoid hand has no glove at all or may wear one? The answers vary by use, and this page gives a way to test them.
What hand-tracking benchmarks assume
The reference datasets in this area record bare hands under controlled conditions. DexYCB records 10 subjects grasping 20 YCB-Video objects from 8 views and benchmarks 3D hand pose estimation, among other tasks (Chao et al., 2021). EgoDex collected 829 hours of egocentric video with paired 3D hand and finger tracking from an Apple Vision Pro, across 194 tabletop tasks with household objects (Hoque et al., 2025). Many estimators use a parametric hand model such as MANO, learned from about 1,000 high-resolution 3D hand scans of 31 subjects (Romero et al., 2017).
None of the facts above say how these methods perform on gloved hands, and this page does not claim to know. The point is a procurement one: a vendor's tracking quality measured on bare-hand benchmarks is not evidence about your footage. Ask for measurements on gloved samples that resemble your tasks.
How gloves and PPE change the problem
| Factor | Effect on hand-pose extraction | Effect on robot transfer |
|---|---|---|
| Skin-colored vs colored gloves | Colored gloves may help segmentation but shift appearance away from training distributions | Appearance gap between human and robot hand grows |
| Bulky gloves | Finger joint angles are harder to infer; contours are blurred | Finger motion is less informative of fine finger use |
| Loose or textured gloves | Wrinkles mimic or hide joints | Estimates may be noisy |
| Sleeves, gauntlets | Wrist location harder to see | Wrist pose errors propagate to end-effector targets |
| Gloves with tools or wires | Occlusion by grips and cables | Hand-object relations partly hidden |
| Other PPE (glasses, hard hats, masks) | Head-mounted rigs may not fit; lens glare; face obstruction | Camera position differs from standard rigs |
Bulk is not purely a disadvantage. Gloves limit finger dexterity, so many gloved tasks involve coarser grasps that are easier to approximate than in-hand manipulation. For tasks like pallet handling, a gloved power grasp may be both common and trackable at the whole-hand level, if not at fingertip precision.
What gloved-hand footage can and cannot support
Can support
- Task and step recognition, which depends on scene and object context more than finger detail.
- Whole-hand and wrist trajectories for coarse tasks.
- Learning how PPE-wearing workers do the task, if the deployment context involves PPE.
- Safety and compliance checks if PPE is the subject.
Cannot support
- Fingertip-accurate 3D hand pose from video alone, unless validated on gloved examples.
- Contact forces and grip pressure.
- Robot hand actions. Retargeting from a gloved human hand to a robot hand is a separate problem; see retargeting human hand demonstrations.
- Fine in-hand manipulation inferred from video; see in-hand manipulation from video.
Decision guide
- If your robot works with bare hands or its own gloves, and you need finger detail: record bare-handed demonstrations under controlled conditions, or use instrumented capture such as gloves with sensors. Gloved footage becomes context, not hand-pose ground truth.
- If your robot must work around gloved humans or in PPE environments: keep the gloved footage; it matches the scenes. Evaluate perception on those scenes.
- If gloves are essential to the task (hazards, hygiene, cold): the demonstration itself requires gloves, and you should plan collection or annotation that accepts them.
- If you are co-training with tracked hands: run a pilot to measure tracking error on gloved samples before purchase. See 3D hand pose data.
Illustrative evaluation set for gloved-hand footage
# Illustrative only - synthetic values; buyer sets the numbers
sample:
clips: 100 stratified
strata: [bare, thin_glove, heavy_glove, sleeve_gauntlet]
tasks: representative of buyer's deployment
labels_for_test:
hand_keypoints: manually annotated 2D keypoints on 20 frames per clip
hand_visibility: per-frame: both | one | none | occluded_by_object
reported_metrics:
tracker_dropout_rate: per stratum
2D_keypoint_error: per stratum, in pixels and as fraction of hand size
wrist_position_error: per stratum
decision_rule: buyer-set tolerance for each stratum
Do not buy at scale until you have run this on the supplier's real footage. Review sample annotations alongside hand-object interaction annotation, and consider the effect of glove color, which can interact with segmentation models.
Collection options for the gloved-hand gap
- Instrumented gloves or sensors under or over PPE, capturing finger motion directly. These add cost, fatigue and sometimes safety questions; confirm that sensors do not alter the task.
- Paired capture: the same task done bare-handed and gloved by the same demonstrator, so you can learn the correspondence.
- Manual annotation of key frames to supplement automated tracking.
- Wrist cameras that view the hand close up; see wrist-mounted cameras.
Whether to stay with existing footage or commission new capture depends on how much of your target use needs finger-level data, as discussed in existing footage vs commissioned recording.
Permissions and provenance
Footage with PPE often comes from regulated settings such as food, health-adjacent services, labs or construction. Ask about site rules for recording, customer or patient presence, and whether safety footage was recorded for incident review rather than reuse. Hand and body measurements derived from video may attract special rules in some jurisdictions; this is not legal advice, and you should ask the supplier for its assessment. Also see safety behaviors in demonstrations.
Questions to ask suppliers
- What share of footage shows gloved hands, and which glove types?
- Has hand tracking been evaluated on gloved samples, and can you share the results?
- Which camera positions are used, and are hands visible during grasp and release?
- Is any footage from sites with regulated hygiene, safety or privacy rules?
- Are bare-hand and gloved takes of the same task available?
- Can annotators mark hand visibility and glove type per clip?
- Was PPE interfering with the camera mount, and are there gaps in coverage?
Discuss your requirements
If your project depends on hands in PPE, describe the glove types, tasks and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match, and SourceX follows up about potential sources. Start from the humanoid training data guide or the workplace task demonstration hub; the catalog entry on egocentric video of manual work describes related footage.
Discuss your humanoid training-data requirements with SourceX
Sources
- EgoDex: Learning Dexterous Manipulation from Large-Scale Egocentric Video — egocentric video with paired 3D hand and finger tracking.
- DexYCB: A Benchmark for Capturing Hand Grasping of Objects — controlled hand-pose benchmark.
- Embodied Hands: Modeling and Capturing Hands and Bodies Together — MANO hand model.