Short answer
Teleoperated demonstrations record the robot's own camera views, joint states and commanded actions, so they can train a control policy directly, but every episode needs a robot, an operator and a reset. Human demonstrations are cheaper to scale and more diverse but contain no robot actions, so they help through pretraining or co-training. Most humanoid programs need both, plus middle routes such as handheld grippers and hand-tracking rigs.
Two supply chains, one budget
These two data types come from different places. Teleoperated episodes come from robots, trained operators and specialist collection teams. Human demonstrations come from people doing the work: existing footage held by operating businesses, or new recordings commissioned at their sites.
Confusing the two is expensive in both directions. Teams buy hours of workplace video when their bottleneck is action data on their own robot, or pay for teleoperation to get visual and scene diversity that human video supplies more cheaply. This page sets out what each record contains and how to split a budget between them.
What each record contains
| Field | Teleoperated robot episode | Human demonstration video | Human demonstration with tracked hands |
|---|---|---|---|
| Images | From the robot's own camera rig | From a head, wrist or fixed camera on or near a person | From headset or glasses cameras |
| Proprioception | Joint positions and velocities | None | Hand, finger and head pose only |
| Actions | Commanded joint or end-effector actions at the control frequency | None | Estimated human hand trajectories, not robot commands |
| Hand or gripper | Commanded gripper or hand state | None | Human finger poses |
| Contact force | Only if the robot has force sensing and it is logged | None | None |
| Scene diversity | Limited by robots, transport and setup | High: any workplace a camera can enter | Moderate: wherever the device can be worn |
| Main cost drivers | Hardware, operators, resets, maintenance, calibration | Recording, consent, rights review, annotation | Devices, calibration, consent, pose processing |
The robot episode is the only one of the three that matches the robot's observation and action spaces by construction. The other two need a bridge, discussed on the human-to-humanoid embodiment gap.
What published teleoperation programs involve
At scale. DROID collected about 76,000 trajectories and 350 hours of interaction across 564 scenes and 86 tasks, using 50 data collectors over 12 months (Khazatsky et al., 2024). Each station used the same rig: a Franka Panda arm, a Robotiq gripper, two exterior stereo cameras and a wrist camera, teleoperated with a Meta Quest 2 controller at 15 Hz, replicated at 13 institutions, with camera calibration, depth and language instructions recorded per episode. That is the shape of a serious teleoperation program: identical hardware, trained people and per-episode metadata.
Small and precise. On low-cost bimanual hardware, Action Chunking with Transformers reached 80–90% success on 6 fine-grained real-world tasks with only 10 minutes' worth of demonstrations (Zhao et al., 2023). Mobile ALOHA added whole-body teleoperation of a mobile base, used 50 demonstrations per task, and found that co-training with existing static bimanual datasets raised success by up to 90% (Fu et al., 2024).
On humanoids. Open-TeleVision used immersive VR with stereoscopic active visual feedback to mirror an operator's arm and hand movements on two different humanoid robots, then deployed imitation-learned policies on long-horizon tasks (Cheng et al., 2024).
Operators matter. In a controlled study of offline learning from demonstrations, results depended on demonstration quality; its multi-operator datasets of 300 demonstrations from 6 operators were grouped by operator proficiency as "worse", "okay" and "better" (Mandlekar et al., 2021). Buy teleoperation data with operator IDs and proficiency information, not as an undifferentiated pool.
Pooling helps but does not replace your own embodiment. Open X-Embodiment combined more than 1 million trajectories from 22 embodiments contributed by 21 institutions (Open X-Embodiment Collaboration, 2023); a new humanoid still needs data on its own body.
What human demonstrations add
Human recording scales without robots. Ego4D gathered 3,670 hours of daily-life video from 931 camera wearers in 74 locations across 9 countries (Grauman et al., 2021), a breadth of people and places that is hard to match with robot collection. When hands are tracked in 3D, human data can be co-trained with robot data in one policy (Kareer et al., 2024), and one humanoid study used roughly 27,000 human demonstrations alongside roughly 1,500 humanoid ones (Qiu et al., 2025). Those ratios belong to those studies; they are not targets.
For the full list of uses, from visual pretraining to latent actions, see what human demonstration video can and cannot do for robot learning.
Middle routes
- Handheld grippers. The Universal Manipulation Interface puts a robot-like gripper in a person's hand, with a wrist-mounted fisheye GoPro as the only sensor and IMU-aided SLAM for tracking; the resulting policies are hardware-agnostic and deployable across robot platforms (Chi et al., 2024). The view and gripper match a robot's wrist, but not a humanoid's body. See handheld gripper demonstrations.
- Hand motion-capture rigs. DexCap records hands and 3D scene observations with a portable rig and learns dexterous robot policies from them, within roughly 40-minute sessions limited by power and with no force sensing (Wang et al., 2024).
- Headset hand tracking. EgoDex recorded 829 hours of egocentric video with paired 3D hand and finger tracking across 194 tabletop tasks using a mixed-reality headset (Hoque et al., 2025).
These routes aim to lower the cost per demonstration relative to teleoperation while keeping more structure than plain video. None of them records the humanoid's joint states.
Choosing a primary source
| If you need | Primary source | Supplement with |
|---|---|---|
| A policy that runs a defined task on your humanoid | Teleoperation on that robot | Human video for scene and object diversity |
| Generalization across sites, objects and lighting | Human demonstrations, existing or commissioned | Teleoperation to ground actions |
| Dexterous finger control | Teleoperation with hand retargeting, or a hand motion-capture rig | 3D hand-tracked human data |
| Contact-rich insertion or fastening | Teleoperation with force logging, or instrumented rigs | Human video only for task structure |
| Step recognition, task planning, progress estimation | Human video with language | None needed for control |
| Evaluation of a deployed policy | Robot rollouts under held-out conditions | Human video to define tasks and success |
If you are unsure whether you are buying for imitation or for pretraining, settle that first; demonstrations for imitation learning vs visual pretraining explains why the two need different data.
Illustrative blended plan
# Illustrative only — tiers, purposes and checks are examples; quantities are set by your own pilot
tier_1_robot:
source: teleoperation on humanoid_X (specialist collection team)
purpose: imitation learning for 6 target tasks; held-out evaluation
key_spec: joint + end-effector actions, 3 cameras, calibration per session, operator_id
acceptance: a random 5% of episodes pass timing, calibration and action-range checks
tier_2_tracked_human:
source: commissioned recording, site_004 and site_011, consenting staff
purpose: co-training on the same task family
key_spec: calibrated head-mounted camera, 3D hand and head pose, task IDs matching tier 1
acceptance: hand tracking present on most frames of each clip; tasks map to tier-1 IDs
tier_3_video:
source: licensed existing footage from operating businesses
purpose: visual and latent-action pretraining
key_spec: hands visible, constant frame rate, step segments, permitted-use terms
acceptance: sample audit before purchase
quantities: set after a pilot measures how each tier moves your evaluation
Where business suppliers fit
Operating businesses can supply the human side: existing footage, and site access with consenting workers for new recordings. They rarely hold teleoperation data for your robot; the exception is a business that already runs robots and logs them. Teleoperation needs robot hardware, trained operators, calibration and episode QA on your embodiment.
SourceX's role is sourcing, rights review, licensing and delivery coordination; it can also arrange annotation and engage specialist capture partners, scoped case by case. For the robot-specific specification, see teleoperation data collection requirements, and for when to license footage versus record it, see existing workplace footage vs commissioned recording.
Limitations on both sides
- Teleoperation inherits the operator's habits, the interface's latency and the robot's current hardware. A new hand or camera mount can make older episodes less useful.
- Teleoperated scenes are usually few, because every site needs a robot and a setup.
- Human demonstrations lack actions and forces, and their camera rarely matches the robot's.
- Mixing ratios reported in papers are specific to their tasks, models and robots. Requirements depend on architecture, task, embodiment and intended use.
Questions to ask suppliers
- Which robot, hand and camera rig produced the episodes, and how close is it to your humanoid?
- What is the action representation (joint or end-effector, absolute or delta) and the control frequency?
- Which teleoperation interface was used, and what latency did operators work with?
- How many operators contributed, how were they trained, and is operator ID logged per episode?
- Are failed, aborted and corrected episodes kept and flagged?
- Is calibration (intrinsics, extrinsics, hand-eye) delivered per session?
- For human data: is the camera calibrated, are hands tracked in 3D, and do task IDs match the robot tasks?
- Who owns the episodes, and can they be licensed for training and commercial deployment?
Discuss your requirements
If your plan mixes robot and human data, tell SourceX which tiers you need and what each must contain. SourceX looks for US businesses that hold suitable human-side material, checks the data and the supplier's licensing permissions, agrees allowed uses in a license and coordinates delivery. It follows up about potential sources after reviewing the request; a request does not guarantee a match. The catalog entry for first-person video of skilled manual work describes the related recording program, and the workplace task demonstration hub and humanoid training data guide map the rest of the section.
Discuss your humanoid training-data requirements with SourceX
Sources
- DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset — scale and setup of a multi-site teleoperation collection.
- Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware — precise bimanual skills from a short teleoperated dataset.
- Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation — whole-body teleoperation and co-training.
- Open-TeleVision: Teleoperation with Immersive Active Visual Feedback — immersive teleoperation of humanoids.
- What Matters in Learning from Offline Human Demonstrations for Robot Manipulation — demonstration quality and operator proficiency.
- Open X-Embodiment: Robotic Learning Datasets and RT-X Models — pooled multi-embodiment robot data.
- Ego4D: Around the World in 3,000 Hours of Egocentric Video — scale and diversity of human video.
- EgoMimic: Scaling Imitation Learning via Egocentric Video — co-training human and robot data.
- Humanoid Policy ~ Human Policy — human demonstrations as cross-embodiment data for humanoids.
- Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots — handheld gripper demonstrations.
- DexCap: Scalable and Portable Mocap Data Collection System for Dexterous Manipulation — portable hand motion capture.
- EgoDex: Learning Dexterous Manipulation from Large-Scale Egocentric Video — egocentric video with 3D hand tracking.