The Data Bottleneck

Language model pretraining runs on text scraped at internet scale. GPT-3 consumed 300 billion tokens in 2020 Brown 2020Language Models are Few-Shot LearnersTom B. Brown, Benjamin Mann, Nick Ryder, NeurIPS 2020, 2020; Llama 3 consumed over 15 trillion in 2024 Meta AI 2024Introducing Meta Llama 3: The most capable openly available LLM to dateMeta AI, 2024, a scale the open FineWeb corpus replicates from 96 Common Crawl snapshots Penedo 2024The FineWeb Datasets: Decanting the Web for the Finest Text Data at ScaleGuilherme Penedo, Hynek Kydlíček, Loubna Ben allal et al., NeurIPS 2024, 2024. Robot learning has no equivalent. The largest aggregated robot corpus, Open X-Embodiment, holds over a million trajectories across 22 robot embodiments Open X-Embodiment Collaboration 2023Open X-Embodiment: Robotic Learning Datasets and RT-X ModelsOpen X-Embodiment Collaboration, 2023, which works out to roughly ten thousand hours of interaction, an estimate rather than a published count. The largest single-organization release, AgiBot World, also claims about a million trajectories, with no published hour count AgiBot Research 2025AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied SystemsAgiBot Research, 2025. The largest open dataset collected on a single robot platform, DROID, holds 76,000 trajectories totaling 350 hours, and collecting those took 50 operators at 13 institutions a full year Khazatsky 2024DROID: A Large-Scale In-The-Wild Robot Manipulation DatasetAlexander Khazatsky, Karl Pertsch, Suraj Nair, 2024. Toyota Research Institute's Large Behavior Model program trained on about 1,700 hours in total, spread across its own bimanual data, simulation, UMI, and Open X-Embodiment TRI LBM Team 2025A Careful Examination of Large Behavior Models for Multitask Dexterous ManipulationTRI LBM Team, Science Robotics 2026, 2025. That gap, and everything the field is doing about it, is the subject of this module.

Two data universes, eight orders of magnitude apart

Text is scraped. The internet accumulates language as a byproduct of everything people publish, so assembling a 15-trillion-token corpus is a filtering problem, not a collection problem. Robot data is different. Every hour of it is produced by a physical machine interacting with the world in real time, usually driven by a paid human teleoperator. There is no crawl for contact.

The chart below plots both universes on one log-log plane. Robot and human demonstration data sits on the bottom lane, read off the hours axis: DROID at 350 hours, TRI's LBM corpus near 1,700, Open X-Embodiment at an estimated 10,000, AgiBot World at an estimated 100,000. Language corpora sit on the left lane, read off the tokens axis: GPT-3 at 300 billion, Llama 3 at 15 trillion. The dashed diagonal spans the empty plane between the two largest entries. No exchange rate between an hour of robot data and a token of text is drawn, because no honest one exists.

The slider projects a teleoperation farm. Under the dedicated-farm assumption, about four productive teleop hours per rig per day, ten rigs accumulate OXE-scale data in a year. Switch to the measured DROID rate, 350 hours from 50 collectors in 12 months, and the same target takes over a century at that fleet size. Real efforts sit between these poles, but the arithmetic explains the industry's response: teleoperation farms, where rows of operators drive fleets of rigs full time. It also gives you the two questions to ask of any data-scale claim. First, how many hours, because episode counts hide trajectory length. Second, how diverse, which is where the scaling laws come in.

What the scaling laws say: diversity beats density

The obvious answer to the bottleneck is to collect more. Two papers now give quantitative guidance on what "more" should mean, and both point away from raw volume.

Lin et al. collected over 40,000 demonstrations and executed more than 15,000 real-world rollouts to map how generalization scales in imitation learning Lin 2024Data Scaling Laws in Imitation Learning for Robotic ManipulationFanqi Lin, Yingdong Hu, Pingyue Sheng et al., ICLR 2025, 2024. Success on unseen environments and objects follows a rough power law in the number of training environments and objects. Demonstrations per environment saturate quickly: past about 50 per environment, adding more yields little, while adding environments keeps paying. Their efficient recipe, 32 environments with 50 demonstrations each, reached about 90% success in novel environments with unseen objects, and the data for it was collected by four operators in a single afternoon.

Shi et al., working on the AgiBot World stack, took diversity apart into three axes: task, embodiment, and expert Shi 2025Is Diversity All You Need for Scalable Robotic Manipulation?Modi Shi, Li Chen, Jin Chen et al., 2025. Task diversity mattered more than per-task demonstration count, echoing Lin et al. Multi-embodiment pretraining turned out to be optional: models pretrained on high-quality single-embodiment data transferred to other platforms and scaled better during fine-tuning than multi-embodiment models. Expert diversity actively hurt. Operators with different styles produce multimodal action distributions, velocity multimodality in particular, that confound imitation. Debiasing those distributions produced GO-1-Pro, a 15% performance gain the authors equate to a 2.5x increase in pretraining data.

The strategic reading is that the bottleneck is breadth, not hours. An hour of teleoperation in a new kitchen with new objects is worth more than an hour in a scene the policy has already seen fifty times. A data budget spent maximizing scene and task coverage beats the same budget spent deepening coverage of familiar setups, and a fleet of operators with identical style can be worth more than a larger fleet with mixed styles, unless the mixed styles are debiased.

Around the bottleneck: data that does not need a robot

If breadth is the constraint, sources that are cheap and broad beat sources that are expensive and narrow, even when they lack robot actions.

UMI removes the robot from collection entirely: an operator holds a 3D-printed gripper fitted with a GoPro and performs the task wherever the task naturally lives, and policies trained on the handheld data deploy zero-shot onto real arms Chi 2024Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild RobotsCheng Chi, Zhenjia Xu, Chuer Pan et al., 2024. Apple's EgoDex used the Vision Pro's hand tracking to record 829 hours of dexterous human manipulation with per-joint 3D poses across 194 tabletop tasks, again with no robot in the loop Hoque 2025EgoDex: Learning Dexterous Manipulation from Large-Scale Egocentric VideoRyan Hoque, Peide Huang, David J. Yoon et al., ICLR 2026, 2025. Ego4D offers 3,670 hours of egocentric daily-life video from 931 camera wearers across 74 locations, valuable for visual pretraining even though it carries no action labels Grauman 2022Ego4D: Around the World in 3,000 Hours of Egocentric VideoKristen Grauman, Andrew Westbury, Eugene Byrne, CVPR 2022, 2022.

EgoScale is the strongest evidence so far that this route scales. Zheng et al. trained a vision-language-action model on 20,854 hours of action-labeled egocentric human video, more than twenty times larger than prior efforts, and uncovered a log-linear scaling law between human data scale and validation loss Zheng 2026EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human DataRuijie Zheng, Dantong Niu, Yuqi Xie et al., 2026. That validation loss, in turn, correlated with real-robot performance, which makes human video a predictable supervision source rather than a hopeful one. Their two-stage recipe, large-scale human pretraining followed by a lightweight aligned human-robot mid-training, improved average success rate by 54% over a no-pretraining baseline on a 22-DoF dexterous hand and transferred to robots with simpler hands.

Simulation is the other escape route: GPU-parallel environments compress months of interaction into hours of wall-clock, at the price of the sim-to-real gap covered in massively parallel sim RL. And whichever source the data comes from, the failure modes of learning from it in closed loop are the ones covered in behavior cloning foundations: covariate shift does not care whether the demonstrations came from a teleop rig, a GoPro, or a physics engine.

How to read the field's data claims

Three habits serve well here. First, convert everything to hours: episode counts hide trajectory length, and a million ten-second episodes is under three thousand hours. Second, ask for the diversity breakdown, meaning environments, objects, tasks, and operator styles, because the scaling laws say those axes dominate the hour count. Third, treat vendor-reported scale as a claim until the data or an independent audit is public; AgiBot World's estimated hundred thousand hours AgiBot Research 2025AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied SystemsAgiBot Research, 2025 is the largest figure on the chart and the least independently verified. The bottleneck is real, but so is the incentive to sound past it.