Cost per hour is how most robotics teams buy data, and cost per hour is the wrong metric. The four layers of the robotics data stack differ by orders of magnitude in what an hour costs. Yet the budgets that fail rarely fail on price. They fail through four predictable leaks that no hourly rate reveals, and the research of 2025 and 2026 documents the mechanism behind each one.
The stack is a pyramid. At the base sits internet-scale human video, which does not by itself provide recorded robot commands or measured robot state. At the top sits real robot data, including teleoperation: a human remotely driving a robot to demonstrate tasks. Its value depends on what is actually recorded. Commands, measured joint state, and force or tactile measurements are distinct channels; none should be inferred from a video. Between them sit simulation and world-model synthetic data, and human demonstrations. Within that last category, ordinary first-person video and instrumented capture such as handheld grippers provide different signals and require different conversion work.
In April 2026, Bessemer Venture Partners forecast that the industry would spend more than $3 billion on robotics data over the following two years. Its estimate covers teleoperation, first-person video, simulation and physical demonstrations.
We have already argued that data, not architecture, now decides who ships. What nobody has written down is how to spend. Consider this the audit: what each layer costs, what it alone can teach, and the four places the money leaks.
The pyramid prices abundance against signal
The models in question, Vision-Language-Action (VLA) policies (models that map camera images and language instructions directly to robot actions) and world models, train on data sources whose prices differ by orders of magnitude, and the gradient tracks the scarcity of embodied signal, not the usefulness of an hour. Bessemer estimates roughly 300,000 hours of robot manipulation data worldwide, compared with about one billion hours of internet video. A lab can draw on video to learn what a task looks like and reserve time on its own robot to learn which actions work.
Internet video can teach a model how objects and people move. Preparing it for training takes work: NVIDIA’s Cosmos paper describes a pipeline for selecting and processing clips before training. Budget for access rights, filtering, compute and storage. Being able to watch a clip does not establish permission to train on it.
Meta pretrained V-JEPA 2 on images and more than a million hours of internet video. The team then trained its action-conditioned model on under 62 hours of DROID robot recordings. It used goal images to plan pick-and-place actions on Franka arms in two labs without collecting training data in those deployment environments. The 62 hours describe one stage of the recipe.
Ego4D illustrates what first-person video can cover: 931 camera wearers recorded 3,670 hours of activity across 74 locations. If your training loader also needs commanded gripper positions or joint measurements, specify how those signals will be recorded or derived.
NVIDIA reports generating synthetic data for GR00T N1.5 in 36 hours. Its estimate of nearly three months for manual collection describes what the company expected that alternative to take. To compare costs on your task, include compute, preparation of the input examples, and time spent reviewing the generated trajectories.
Generating motion and extracting actions can have separate compute costs. For example, 1X reported that generating five seconds of video took 11 seconds on multiple GPUs. Extracting the actions from that video took another second.
An instrumented gripper lets an operator demonstrate a motion without occupying the target robot. In DexViTac’s four-task experiment, collection averaged 14.5 seconds per demonstration with its handheld system. The Universal Manipulation Interface study compared a handheld gripper with joystick-based robot teleoperation on cup arrangement. For a pilot, time the complete shift: setup, demonstration, reset and review. Divide the number of accepted episodes by that time to get a rate you can use in a budget.
The 1X training report shows how different stages consume different data. After pretraining on web video, the team used 900 hours of first-person human footage and then 70 hours of robot data to adapt the video model. A separate model learned to infer robot actions from 400 hours of robot recordings. The report does not specify the overlap between the robot datasets, so it gives no unique-hour total for the system. Our world-model analysis explains what each stage contributes.
EgoScale, from NVIDIA's GEAR lab (Generalist Embodied Agent Research), trained a VLA on 20,854 hours of action-labeled egocentric video and reports a "log-linear scaling law between human data scale and validation loss" that "strongly correlates with downstream real-robot performance," improving average success 54% over a no-pretraining baseline on a 22-degree-of-freedom hand (Zheng et al., arXiv 2602.16710, February 2026).
Real robot collection adds embodiment-specific hardware and operating costs. ALOHA and Mobile ALOHA describe research teleoperation systems. Their published designs are useful planning references, but a hardware bill from a paper is not a current supplier quote or a complete collection-program budget. Obtain configuration-specific costs, including maintenance, calibration, and setup.
On top of hardware sit operator time, scene setup, resets, review, and re-recording. A throughput figure is useful only with its task duration, operating conditions, and acceptance criteria. Measure accepted output over total operator time in a representative pilot rather than treating a paper's task-specific rate as a universal ceiling.
The price gradient is real. It is also the least useful fact about the pyramid.
Cost per hour is the wrong denominator
What an hour teaches depends on the task, model, and existing training set. Diversity can be a better use of the next collection budget than repeating conditions already well covered. Lin et al., "Data Scaling Laws in Imitation Learning for Robotic Manipulation" (ICLR 2025; arXiv 2410.18647), found that generalization follows a roughly power-law relationship with the number of environments and objects, while "the diversity of environments and objects is far more important than the absolute number of demonstrations." In the settings they tested, additional demonstrations had diminishing returns after a threshold. The study does not establish one stopping threshold for every robot, task, or model.
For two equally priced shifts, ask what each adds. One might cover a familiar grasp under a new shelf height; another might repeat a known success hundreds of times. A held-out test on that shelf gives you a way to judge which collection was worth buying.
Leak one: paying for repetition the model ignores
The first leak is redundant easy-case teleoperation that starves the long tail (the vast set of rare situations that individually appear almost never and collectively dominate deployment). Lin et al. motivate checking whether additional demonstrations in a familiar setting still improve generalization before spending the next collection budget there. That is a measurement to make on the target task, not a guarantee that repetition has no value.
The DROID paper reports that 50 collectors recorded 76,000 trajectories across 564 scenes and 84 tasks over a year. The 350 recorded hours describe the dataset duration. For a collection budget, the year of fieldwork also matters. The π0.5 paper describes a different collection choice: the authors combined data from multiple robots with web and other training data, then tested mobile manipulation in homes outside the training set. For a household robot, holding out entire homes tests an important part of the deployment brief.
Use held-out evaluation to compare the next collection options: more demonstrations in existing conditions, new objects or environments, and targeted failures or recoveries. Keep the training recipe and evaluation conditions comparable. A plateau may reflect coverage, label quality, optimization, or model limitations; demonstration count alone does not diagnose the cause.
Leak two: simulation compute without a real anchor
Simulation lets a team vary conditions that would be slow to stage repeatedly. In the robotic Rubik’s cube study, the researchers varied simulated conditions during training and evaluated the resulting policy on a physical robot hand. A buying brief should identify which part of the task the simulator represents well enough to vary.
Consider a gripper lifting a wet container. You can render many views of the container while leaving the contact model identical in every example. Reserve physical trials that vary the wetness and grip, record where the object slips, and compare the simulated and observed outcomes. If the policy improves in simulation while the same physical failures persist, investigate that mismatch before paying for another generated batch. See our analysis of the sim-to-real gap.
Leak three: datasets depreciate, loops compound
When a shelf moves or a gripper changes, which recordings still teach the task? An old camera view might still show where to reach. Its recorded grasp command may fit only the old gripper.
Some experience travels across robots. The Open X-Embodiment collaboration assembled more than a million trajectories across 22 robot embodiments. In tests at five labs, RT-1-X had a mean task success rate of 63%. The original methods trained on each robot’s own dataset averaged 41%. Before buying pooled recordings, check which observations and actions map to your robot and test that transfer on your task.
The value of pooling depends on the task and the measure. In the 1X study, human reviewers sometimes scored generated videos lower after first-person data was added for familiar tasks. In physical trials, the variants performed similarly on the familiar task of grabbing chips. On dish scrubbing, only the variant trained with both first-person footage and enriched captions succeeded at least once. The authors ran each variant 30 times per task. For a buying decision, measure success on the target robot as well as the quality of generated video.
Take one failed grasp from the fleet and trace it through the next release. Can the team find the original recording, explain the review decision and rerun the test after retraining? That record helps a buyer decide whether to collect another example, revise a label or change the test. How to plan a collection.
Leak four: testing only familiar conditions
The first hidden cost is a test set that never varies the camera, lighting or starting position. In Table 1 of LIBERO-Plus, OpenVLA completed 76.5% of tasks on the original simulated benchmark and 1.1% under camera perturbations. A different model, OpenVLA-OFT, scored 97.1% and 59.7% in those same columns. Test the camera changes your robot will encounter and report the results for the model you intend to deploy.
The second hidden cost is finding unusable episodes after delivery. Ask reviewers to inspect a representative slice and report why they rejected each episode. A corrupt timestamp calls for a different fix from a missing camera angle or an ambiguous task instruction. Budget for the corrective work as well as the initial recording.
For high-risk AI systems, Article 10 of the EU AI Act addresses where data came from, how it was prepared, and whether it represents the conditions in which the system will be used.
Put those questions in the collection brief. If a robot will pick parcels in a dim stockroom, ask how many evaluation episodes cover that lighting, who checked them, and where the gaps are. Include the cost of collecting and reviewing the missing cases in the budget. Those are the cases a team needs to test before it puts the robot to work.
The strongest objection: fleets will make collection free
If deployed robots already record the work, why pay for more collection? A fleet can reveal conditions a lab missed. Its logs can also show whether a problem is common enough to justify collecting more examples.
Waymo and Swiss Re compared liability claims over 25.3 million fully autonomous miles with a human-driver benchmark. They reported nine property-damage claims and two bodily-injury claims for the Waymo fleet, against 78 and 26 expected at the benchmark rates. That measures a fleet outcome. The comparison does not isolate how much any collection or labeling decision contributed. Read the study summary.
For a buyer, the first question is what the fleet already sees. If you have hundreds of recordings of a failed grasp, budget for reviewing the failure and testing the fix. If the gripper blocks the camera at contact in every recording, collecting more of the same view leaves the cause hidden. That is a reason to change the capture setup.
Budget by what each layer alone can do
Start with one target task, the signals it requires and a test in a held-out environment. For each data source, specify what it contributes: video for visual context, simulation for a condition you can vary, instrumented demonstrations for motion, or robot recordings for the commands and state your policy needs.
Cost per accepted episode is the total cost of setup, capture, compute, review and rework divided by the number of episodes that pass the agreed checks. Count failures against the collection budget even when you exclude them from delivery. A failed grasp with a usable recording and a labeled recovery can still pass if recovery is part of the brief.
Compare sources on the same output. If one delivers video and another delivers synchronized observations and actions, put the cost of producing the missing signals into the comparison first. Then decide whether the next dollar belongs in more collection, better review, or a stronger deployment test.
Related: how a teleoperation session moves through a 4D pipeline and what VLA models need from their training data.



