Every scaling-law conversation in robotics eventually borrows its vocabulary from language models, and every time it does, the analogy quietly breaks in the same place. GPT-style models scaled because the training data already existed — Common Crawl, GitHub, Books3, a decade of the internet writing itself down for free. Robot manipulation policies have no equivalent corpus. Open X-Embodiment, the largest cross-institution effort to date, aggregates on the order of a million episodes across 20+ labs and 20-some embodiments. That sounds like scale until you compare it to what it's standing in for: a few thousand hours of largely tabletop pick-and-place, lab lighting, a narrow band of grippers, and demonstration policies that were often optimized for "does this work" rather than "is this representative." It is not the internet. It's closer to a good university's worth of lab notebooks.
That gap matters because the empirical results we do have point somewhere uncomfortable. RT-X, Octo, and the follow-on scaling studies on manipulation data agree on one thing more than any other: task and embodiment diversity moves policy performance more reliably than raw episode count. A few hundred more demonstrations of the same bin-picking task on the same arm buys you very little past a point of diminishing returns. A few hundred demonstrations spread across different grippers, camera placements, object geometries, and failure recoveries buys you generalization. This is the opposite of what the LLM playbook trained us to expect, where volume was mostly the answer and diversity came along for free because the internet is naturally diverse. In robotics, volume and diversity are in tension, because every additional hour of data has to be physically generated by a human moving a physical arm.
The naive fix, and why it doesn't scale
The obvious response is to treat this like any other data problem: hire more teleoperators, build more rigs, run more shifts. It's the instinct that made sense for data labeling — throw more annotators at ImageNet, throw more Mechanical Turk workers at RLHF preference pairs. Robot demonstration collection doesn't behave the same way, for a simple reason: a labeler classifying an image is fungible in a way an operator producing a force-controlled insertion demonstration is not.
Good teleoperated demonstrations are a skill, not a task. A new operator on a 6-DoF arm with a parallel gripper needs real hours before their trajectories stop looking like exactly what you don't want a policy to imitate — overcorrection near contact, inconsistent grasp approach angles, hesitation that shows up as multimodal action distributions in the training set. Getting an operator to the point where their demonstrations are interchangeable with your best operator's, on a contact-rich task like connector mating or deformable-bag handling, typically takes weeks of guided practice, not a single onboarding session. Compress that ramp and you don't get more data faster — you get more data that quietly degrades whatever policy trains on it, because imitation learning methods (ACT, diffusion policies, and their descendants) reproduce operator noise as faithfully as they reproduce operator skill.
Hiring at the naive rate also runs into a geography wall before it runs into a skill wall. If demonstration collection requires an operator physically present at the rig, your labor pool is whoever lives within commuting distance of your lab — a few hundred people in most cities, fewer if you need any manipulation experience at all.
Where the leverage actually is
None of this means the data bottleneck is unsolvable — it means the solutions look like labor-market and workflow engineering more than they look like storage engineering.
- Remote teleoperation widens the labor pool past lab geography. Once glass-to-glass latency is low enough for contact-rich control (sub-100ms, not the 300-500ms that turns manipulation into "move and wait"), an operator's location stops mattering. Your candidate pool goes from "commutable to this building" to "anyone with a stable connection," which changes hiring economics by an order of magnitude and lets you recruit for manipulation aptitude instead of proximity.
- Fleet supervision multiplies robots per operator instead of operators per robot. A skilled operator supervising several robots executing a learned policy, intervening only on the tail of hard cases, produces far more usable hours per person-week than one operator fully teleoperating one arm. The interventions themselves are disproportionately valuable data — they're concentrated exactly on the distribution edge where the current policy is weak, which is the data a scaling-law-optimal collection strategy would ask for even if you designed it by hand.
- Every session has to land as training-ready data with zero plumbing tax. If an operator's hour of work turns into an hour of engineer time reformatting bag files, re-syncing timestamps, or hand-labeling episode boundaries before it's usable, you've just halved your effective throughput and moved the bottleneck from the rig to the data team. Capture that translates directly into LeRobot, RLDS/TFDS, or HDF5 with synchronized multi-camera streams and episode metadata already attached removes a second, invisible people-bottleneck sitting downstream of the first one.
- Skill curves and QA loops are collection infrastructure, not HR overhead. Tracking per-operator success rate, intervention frequency, and trajectory smoothness over time — and feeding it back as targeted practice rather than generic onboarding — is what turns "weeks to competence" into something you can actually forecast and plan a collection campaign around.
Why simulation and video don't make this go away
Simulation and internet video pretraining are genuinely useful, and neither removes the real-robot data requirement. Sim gives you volume and controllability but still carries a sim-to-real gap in contact dynamics, sensor noise, and lighting that no amount of domain randomization has fully closed for force-sensitive manipulation — you still need real demonstrations to calibrate and correct what the simulator gets wrong. Video pretraining (learning visual and possibly dynamics representations from web-scale footage) gives a policy a better prior over "what the world looks like," but web video has no paired action labels and essentially no force-torque signal, so it can warm-start a policy's perception stack without ever teaching it what torque profile a successful insertion feels like. Both are accelerants. Neither is a substitute.
The scarce commodity
If diversity beats volume, and volume is bounded by operators rather than by disk space, then the resource every lab and every robotics company is actually competing for is operator-hours that produce training-ready diversity — not headcount, not GPUs, not another export format. That's the constraint we built Telemanual's teleoperation and capture pipeline around: widen who can operate through low-latency remote control, multiply what one skilled operator can supervise through fleet oversight, and remove the data-plumbing tax so every session — teleoperated or supervised — turns into a labeled, versioned demonstration without an engineer touching it in between.