The four planned configurations add up to the announced total, line by line. EgoStandard, which uses head-mounted cameras and carries 3D hand-pose annotations, is planned at 80,000 hours with hand poses plus another 10,000 with hand and full-body poses [5]. EgoPro is planned at 8,000 hours of synchronized head-and-wrist footage with hand poses plus 2,000 with hand and full-body poses [6]. 80,000 + 10,000 + 8,000 + 2,000 = 100,000 [1]. That is a collection plan with parts, which is more than most dataset roadmaps offer. Where the announcement gets loose is granularity: it counts tasks and scenes in the thousands [2] while listing 128 scene types and 18 task categories [10], and it does not map one onto the other. It also does not break out how the released 10,000 hours distribute across the four annotated configurations, which is what the per-subset dataset cards carry [9]. EgoDemo, a 50-hour sample drawn from all four annotated configurations plus two raw-video variants, is the cheap way to inspect that before pulling anything large [8].
The wrist camera is the mechanism to read closely. Head-mounted video fails in one specific place: hands leave the frame or obscure the object at the moment of contact [6]. That is the frame a manipulation policy most needs labelled. EgoPro exists to cover it, and its planned share is 10,000 hours against EgoStandard's 90,000 [2], putting the contact-visible configuration at ten percent of the corpus [3]. If your objective concentrates loss on grasp and release, ninety percent of the plan uses the view with the known occlusion weakness [3]. If you are pretraining a general visual representation, that ratio matters much less.
On scale, the tranche that exists is already the largest thing in its neighbourhood. EgoVerse, the consortium Lightwheel belongs to, publishes a snapshot of 4,003 hours, 1,965 tasks and 240 scenes [12]. Divide it out: the released hours are roughly 2.5 times that snapshot, and the full plan would be roughly 25 times it [5]. The release also commits Lightwheel to a collection and annotation operation nine times larger than what it has published [4], and doing that in stages beats annotating 90,000 hours before learning which slices anyone loads. Steve Xie came to this from autonomous-driving simulation at NVIDIA, Cruise and NIO, with a physics degree from Peking University and a doctorate in quantitative finance from Columbia [14]. Reasonable preparation for asking a field to value an option with no expiry date.
The pretraining case carries a condition that no hours count satisfies. The framing in runtimewire's report of the Hugging Face announcement is that human video can reduce robotics' dependence on expensive robot-operated demonstrations, while usable transfer still requires robot-specific data [13]. So this lowers the cost of the representation, not the cost of the demonstrations, and the teleoperation line in your budget survives. The batch is described as a corpus for testing pretraining, representation learning and human-to-robot transfer [16], and for a result reported on it to move to your robot, your task distribution has to intersect those 18 categories [10] and the annotated hand poses have to describe motions your gripper can approximate. Both of those you can check from the sample and the cards, and not from a headline number.
