Robots learn from people: cameras on the head, on the wrist or watching the work, motion capture and teleoperation. Which route works, what it costs, who supplies it and who argues what, with a source for every fact.
Robots need data that the internet does not have: what hands do, with which force, in which order. In 2026 the field turned to people as the source. Cameras on the head (smart glasses, helmets, phones worn on the forehead), cameras on hand-held grippers and gloves, cameras that watch the work from the side, and motion-capture suits now record human work at a scale teleoperation never reached. NVIDIA's EgoScale trained on 20,854 hours of first-person human video and found a log-linear scaling law (arXiv 2602.16710); Build AI published 100,405 hours of factory-worker video (Hugging Face); Generalist says its GEN-1 model was pretrained on more than 500,000 hours from low-cost wearables with no robot data (Generalist, 2 Apr 2026); and JD.com aims for 10 million hours in two years (Rest of World, 3 Jun 2026). A new supply chain followed: data factories, gig apps that pay people to film their chores, and makers of capture rigs. The open questions are how much robot data is still needed, which camera position works, how to close the gap between a human body and a robot body, what an hour really costs, who owns the footage of the people filmed, and whether video without touch is enough.
At equal volume, first-person human video beat teleoperated robot data for pretraining in HumanScale: 5,000 hours of each, with 24% lower validation loss and higher success in and out of distribution, as reported by the authors (arXiv 2606.20521).
Human video scales predictably: EgoScale reports a log-linear law between hours of human video and action-prediction loss, and a 54% gain on a 22-degree-of-freedom hand (arXiv 2602.16710).
One hour of human hand data was worth more than one extra hour of robot data in EgoMimic, which co-trains Project Aria glasses video with robot data (arXiv 2410.24221).
Head and glasses: Ego4D set the base with 3,670 hours from 931 wearers (arXiv 2110.07058); Apple's EgoDex used Vision Pro headsets for 829 hours with tracked finger joints (arXiv 2505.11709); Build AI records factory workers with a head camera (Hugging Face); Meta's Aria Gen 2 research glasses weigh 75 g and track eyes and hands (Road to VR, 27 Feb 2025).
Many cameras on one person: Tesla moved Optimus training in June 2025 from motion-capture suits and teleoperation to video, with workers wearing five cameras on a helmet and backpack, according to Business Insider (via AOL, 25 Aug 2025); Instawork's kit has a stereo head camera, a chest camera and two wrist cameras (Instawork).
Hand-held and wrist: UMI puts a fisheye camera on a hand-held gripper, so policies transfer to any robot with the same gripper (arXiv 2402.10329); DexWild adds palm cameras to a glove and reports 68.5% success in unseen environments (arXiv 2505.07813).
Match the hardware: Sunday designed its robot hand to copy the geometry of its Skill Capture Glove, and says glove data converts to robot data at 90% success (Sunday, 19 Nov 2025).
Paint it away: DexUMI puts an exoskeleton on the hand and paints the robot hand into the video, with 86% average success on two robot hands (arXiv 2505.21864); Phantom does the same with a rendered robot arm (arXiv 2503.00779).
Translate the motion: HumanPlus trained a humanoid's low-level control on 40 hours of human motion and copies a person from one camera (arXiv 2406.10454); Humanoid-X mined more than 20 million poses from internet video (arXiv 2412.14172).
Teleoperation: a bimanual ALOHA station costs about $20,000 (arXiv 2304.13705); a vendor benchmark puts bimanual teleoperation at $28 to $60 per raw hour, against $15 to $40 for first-person video and $20 to $50 for multi-view capture (DexSet, a data vendor, 2 Jul 2026).
Wearables: AirExo-2 costs about $600 per system against about $60,000 for a robot teleoperation platform (arXiv 2503.03081); Sunday's glove is reported at about $200 against $20,000 of teleoperation equipment (secondary report, The Neuron).
Pay for the people filmed: micro1 pays recorders in Nigeria $15 an hour and its CEO says robotics firms spend more than $100 million a year on such data (MIT Technology Review, 1 Apr 2026); home capture in Shandong pays 20 yuan (about $3) an hour (Rest of World, 3 Jun 2026); Indian chore recorders earn about 250 rupees an hour (AFP via Malay Mail, 11 Jun 2026); Tesla paid $25.25 to $48 an hour for motion-capture operators in 2024 (Fortune, 19 Aug 2024).
Scale brings people into the frame: JD.com plans 10 million hours with 100,000 employees and 500,000 outside collectors, with head cameras also used on farms and in elder care (Rest of World, 3 Jun 2026).
Indian factory and service workers are filmed for foreign robot companies through local firms (Scroll, 20 May 2026; Inc42, 26 May 2026); Human Archive raised $8.2 million for this model (same Inc42 source).
Some footage is published openly: Build AI releases its factory video under the Apache 2.0 licence (Hugging Face).
Video does not record force: a 2 N and a 20 N grip look the same on camera, an argument made on the map's robot hands page (focus page).
Rigs are adding touch: Human Archive's kit includes tactile gloves, wrist cameras and motion sensors (Inc42); DexUMI's exoskeleton lets the wearer feel contact, and its policy without touch input failed to grasp a salt shaker (project page); Kaiwang Data combines bare-hand video, tactile gloves and grippers (TechNode, 5 Aug 2026).
The Physical AI Voices by Herbert Scale Experts.