The data bottleneck is the shortage of robot training data compared with the text language models learn from. The largest public robot datasets hold hundreds to a few thousand hoursSource 5 Source 6, while Llama 4 Scout trained on about 40 trillion tokensSource 20 and companies report private corpora of 100,000 hours or more from handheld and wearable devicesSource 21 Source 22.

A grid of robot camera frames grouped by scene type, showing the same Franka arm working in kitchens, bedrooms, bathrooms, offices, laundry rooms and other real buildings.
Frames from DROID grouped into its 9 scene types. The same Franka setup recorded every trajectory, in kitchens, bedrooms, offices and other real buildings.
Paper figure: Alexander Khazatsky et al. / arXiv. Licence: CC BY 4.0.

DROID reports 76,000 successful trajectories totaling 350 interaction hours. Fifty data collectors used 18 robots across 13 institutions over 12 months, and the release adds roughly 16,000 unsuccessful trajectories outside that countSource 5. Drag "Robots collecting" to change fleet size at two authored hypothetical rates: 1,000 or 7 productive hours per rig-year.

Guess first

In an authored hypothetical, each collection rig yields 7 productive hours a year. With 10 rigs, roughly how long until they reach a target of 10,000 hours?
10 part-time robots: 70 hours a year, 143 years to 10,000 hours. Switch to the full-time farm and it takes one year.
The data bottleneck
Robot training data is tiny next to chatbot training text
How much each robot collects
Adjust more
Hours to collect
Datasets to pick out
Chatbot training text, in word piecesGPT-3: 300 billionLlama 3: 15 trillion100 billion10 trillionRobot practice data, in hoursDROID350EgoDex829TRI LBMabout 1,700AgiBot World2,976Ego4D3,670EgoScale20,854Xiaomiover 100,000GEN-0over 270,000GEN-1over 500,0001001,00010,0001 millionHow long to collect 10,000 hours?goal: 10,000 hours101,000100,000hours collected in one year70 hours a year10 part-time robots: 143 yearsto reach 10,000 hours

10 part-time robots collect 70 hours a year: 10,000 hours takes 143 years, 1,000,000 hours takes 14,286 years

robot datahuman videocompany total, not releasedyour robots, one year

Illustrative: chosen collection rates, not a measured farm.

Text piles up online for free; robot data needs people and hardware. The three largest totals are company reports, with no data released.
How this was made

The two panels use separate rulers because there is no honest exchange rate between an hour of robot practice and a token of text, so none is drawn. A token is a piece of a word: GPT-3 trained on 300B tokens and Llama 3 on 15T tokens. Both rulers are logarithmic: each labelled step is ten times the one before.

Authored hypothetical: 7 productive hours per rig-year, not a measured DROID productivity rate. DROID reports collectors and elapsed collection time, not annual exposure for each rig. The 10,000-hour and 1,000,000-hour targets are authored hypothetical inputs, not OXE totals or measured frontier requirements. Hours marked “about” on dataset points are estimates. Hours marked “over” are totals Xiaomi and Generalist report for their own models; neither company has released the data, so no one outside can count them.

  • Full-time farm: about 4 hours a day each (Dedicated farm hypothetical, 1,000 h/rig/yr): Authored hypothetical: 1,000 productive hours per rig-year. For example, 4 hours on each of 250 working days. This is a teaching input, not measured farm productivity.
  • Part-time: 7 hours a year each (Low-rate hypothetical, 7 h/rig/yr): Authored hypothetical: 7 productive hours per rig-year, not a measured DROID productivity rate. DROID reports collectors and elapsed collection time, not annual exposure for each rig.

Current projection: 10 rigs at 70 h/yr, 10,000 h in 143 yr and 1,000,000 h in 14,286 yr. No OXE hour estimate is supplied or plotted here.

Sources for each count: DROID, EgoDex, TRI LBM, AgiBot World, Ego4D, EgoScale, Xiaomi, GEN-0, GEN-1, GPT-3, Llama 3.

Demonstration hours span 350 h (DROID) to 20,854 h (EgoScale) across 6 robot and human datasets, and companies report unreleased totals of 100,000+ h (Xiaomi), 270,000+ h (GEN-0) and 500,000+ h (GEN-1), while pretraining tokens span 300B (GPT-3) to 15T (Llama 3), on separate rulers with no honest hour-to-token exchange rate between them; your 10-rig farm at the low-rate hypothetical rate projects 70 h per year, reaching the authored 10,000-hour target in 143 yr and the authored 1,000,000-hour target in 14,286 yr. No OXE hour estimate is supplied or plotted here.

Datasets and corpora by scale, with the farm projection
datasetdemonstration hourspretraining tokens
DROID350 hn/a
EgoDex829 hn/a
TRI LBM~1,700 hn/a
AgiBot World2,976 hn/a
Ego4D3,670 hn/a
EgoScale20,854 hn/a
OXENo hour estimate plottedn/a
GPT-3n/a300B tokens
Llama 3n/a15T tokens
your hypothetical farm (10 rigs)70 h/yrn/a
OXE: over 1 million recorded robot runs on 22 kinds of robot, with no total hours. This chart supplies no hour estimate and does not plot OXE on the hours axis.
Read the reasoning
  • Around two decades, which needs roughly 500 farm-hours a year instead of 70Twenty years would require 500 hours a year across the whole farm. This scenario produces only 10 times 7, or 70 hours a year, so it takes much longer.
  • Only a few years, on the assumption that each rig collects hundreds of hours a yearThat assumes a higher collection rate than the question supplies. The dedicated-farm hypothetical uses 1,000 hours per rig-year; at 10 rigs, it reaches 10,000 hours in one year.
  • Well over a century, on the order of 140 years at that throughputThe arithmetic is 10,000 hours divided by 70 hours a year, roughly 143 years. Both the rate and target are authored hypothetical inputs. Neither measures DROID productivity or OXE duration.the hours axis

For the same 10,000-hour hypothetical target, 10 rigs take about 143 years at 7 hours per rig-year and one year at 1,000 hours per rig-year. The result depends on the assumed rate.

15T
language
tokens, Llama 3 pretraining
350
robots
interaction hours, DROID dataset
32x50
scaling law
envs x demos, Lin et al.
20,854
human video
hours, EgoScale

Two data universes, nine orders of magnitude apart

GPT-3 trained on 300 billion tokens in 2020Source 1 and Llama 3 on over 15 trillion in 2024Source 2, a scale the open FineWeb corpus rebuilds from 96 Common Crawl snapshotsSource 3. Qwen3 trained on 36 trillion tokensSource 23. Ken Goldberg of UC Berkeley calls the difference a 100,000-year data gapSource 24.

Open X-Embodiment pools over a million trajectories across 22 robot embodimentsSource 4. The chart assigns it no hour estimate. Robot data has no crawl: a paid teleoperation operator usually drives each hour. AgiBot World publishes 2,976 hours for 1,001,552 trajectories, about 11 seconds eachSource 6.

Toyota Research Institute's Large Behavior Models trained on about 1,700 hours of its own bimanual data, simulation, UMI and Open X-EmbodimentSource 7. π0 pretrained on 10,000 hours of dexterous manipulation from 7 robot configurations, plus Open X-Embodiment, DROID and BridgeSource 25.

What the scaling laws say: diversity beats density

Scaling laws come from language modelling: loss falls as a power law in model size, dataset size and computeSource 26. A meta-analysis of 327 robotics papers finds power laws there tooSource 27.

Lin et al. collected over 40,000 demonstrations and ran over 15,000 real-world rollouts on imitation learningSource 8. Success with new environments and objects follows a rough power law in the number of training environments and objects; demonstrations per environment saturate past about 50. Their recipe, 32 environments with 50 demonstrations each, reached about 90% success with unseen objects; four operators collected it in one afternoon.

Eight round fisheye views from a gripper's camera, each a different office or meeting-room table with a mug or bottle in front of the gripper's purple and white jaws.
The eight test environments in Lin and colleagues' data-scaling study, seen through the gripper's fisheye camera. None was in the training data, and every task was tested in all eight.
Paper figure: Fanqi Lin et al. / arXiv. Licence: CC BY 4.0.

On the AgiBot World stack, Shi et al. split diversity into task, embodiment and expertSource 9. Task diversity mattered more than demonstrations per task. Models pretrained on high-quality single-embodiment data transferred to other robots and scaled better in fine-tuning than multi-embodiment models.

Expert diversity hurt: different operator styles make velocities multimodal. Debiasing them gave GO-1-Pro a 15% gain, which the authors equate to 2.5 times more pretraining dataSource 9.

DimensionEffect on performanceSaturationSource
Environmentsstrong positive, power lawnone found up to 32 envsLin et al. 2024
Objectsstrong positive, power lawnone foundLin et al. 2024
Demos per environmentpositive, saturatingabout 50 per envLin et al. 2024
Task diversitystrong positivenone foundShi et al. 2025
Expert diversitynegative unless debiasedmultimodal action distributionsShi et al. 2025
Multi-embodiment pretrainingoptionalsingle-embodiment sufficesShi et al. 2025

Around the bottleneck: data that does not need a robot

UMI collects without a robot: an operator holds a 3D-printed gripper with a GoPro, and its policies deploy zero-shot on real armsSource 10. Apple's EgoDex used Vision Pro hand tracking to record 829 hours of dexterous manipulation with per-joint 3D poses across 194 tasksSource 11. Ego4D holds 3,670 hours of egocentric daily-life video from 931 wearers in 74 locations, without action labelsSource 13. Build AI's Egocentric-100K holds 100,405 hours of head-mounted video of manual labourSource 28.

EgoScale trained a vision-language-action model on 20,854 hours of action-labelled egocentric human video, over twenty times prior efforts. It found a log-linear scaling law between human data scale and validation loss, which correlated with real-robot performanceSource 12. Human pretraining followed by lightweight aligned human-robot mid-training raised average success rate by 54% over a no-pretraining baseline on a 22-DoF dexterous hand. EgoMimic found an extra hour of hand data worth more than an extra hour of robot dataSource 29.

FACT, from Zhi Cao, Jiajun Wu and colleagues at Stanford, the University of Michigan and Microsoft Research (October 2026), was pretrained mostly on about 2,000 hours of human demonstrations recorded with tactile gloves. Fine-tuned on each task, it averaged 41.4% success on seven tasks with two dexterous hands, against 8.6% and 1.4% for T-Rex and π0.5 fine-tuned on the same data. Pretrained without the gloves' force readings, it opened a test-tube cap in none of ten trials instead of eightSource 37.

HiPHI (CoRL 2026) is a motion-capture dataset from Jiahao Ji, Lei Han and colleagues at Noitom Robotics and several universities: 617.5 hours from 132 performers, 245.7 of them with 40 tracked real objects. Retargeted to a Unitree G1, its motions gave the highest simulated tracking success at a matched 3-hour budget against AMASS, LAFAN1, Motion-X++ and BONES-SEED. They led again at 20 hours, a budget LAFAN1 is too short to fill, and policies trained on them ran, crawled and carried a box on a real G1Source 19.

Company corpora reach half a million hours

Generalist AI reports over 270,000 hours of manipulation data for GEN-0, growing by 10,000 hours a week; post-training results improve predictably with more pretraining data and computeSource 30. For GEN-1 it reports over half a million hours from low-cost wearable devices on humans, with no robot data in the base modelSource 22. Xiaomi-Robotics-1 pretrained on over 100,000 hours of trajectories from UMI devices and improved with data and model scaleSource 21.

Sunday trains ACT-1 on glove recordings, converted to robot data with a reported 90% success rateSource 31. Figure reports zero-shot navigation after training only on egocentric human videoSource 32.

Generated and simulated data

GRAIL, from Tianyi Xie, Yuke Zhu and colleagues at NVIDIA and UCLA (June 2026), built over 20,000 humanoid interaction sequences from 3D assets and video models. Policies trained only on them reached 84% pick-up success on a Unitree G1 (five objects, ten trials each) and 90% on stairsSource 14. HumanoidMimicGen, from Kevin Lin, Ajay Mandlekar and colleagues at NVIDIA and UT Austin (May 2026), adapts a few demonstrations to new object poses. Co-training on its data raised average success on four real tasks from 0.51 to 0.71Source 15.

PRISM, from Zihan Wang, Pieter Abbeel, Guanya Shi, Angjoo Kanazawa and colleagues at Amazon FAR, UC Berkeley and Carnegie Mellon (September 2026), turned a few real videos into hundreds of generated variants. Its policy ran on a real humanoid with no real-world fine-tuningSource 16. DexAgent, from Li Fei-Fei, Jiajun Wu and colleagues at Stanford and Columbia (September 2026), turns one egocentric human video into simulated trajectories. Across eleven real tasks, its policies reached 3.5 times the success rate of baselinesSource 17.

MobileVISTA, from Suzannah Wistreich, Jiajun Wu and colleagues at Stanford and the Toyota Research Institute (October 2026), synthesizes images and actions for shifted base poses from demonstrations recorded at one pose. With 50 demonstrations per task and base offsets within 15 cm, its policy stacked a book in 13 of 25 trials, against at most 2 for Canonical-2D, DP3 and a fine-tuned π0.5. At the original pose it stacked 18 of 25, below their 24 and 25Source 18.

Dexterity from Supervised Play, from Botao He, Jitendra Malik and colleagues at Amazon FAR and the University of Maryland (October 2026), starts from 15 minutes of human play on a 22-joint hand. A policy cloned from that play collects data for a world model, then proposes actions for model-predictive control. Asked to rotate a cube to a goal orientation in simulation, the system came within 0.6 radians in 91.4% of 128 trials, where PPO needed about 490 times the data to reach 86.7%, and on the real hand it reached 90.0%Source 38.

Simulation compresses months of interaction into hours but brings the sim-to-real gap covered in massively parallel sim RL. GR00T N1 generated simulated trajectories equal to 6,500 hours of human demonstration in 11 hoursSource 33, and co-training with simulation raised real-world performance by 38% on averageSource 34. A model pretrained only on the synthetic InternData-A1 matched the official π0 on 49 simulated, 5 real-world and 4 long-horizon dexterous tasksSource 35.

Sergey Levine of Physical Intelligence argued in July 2025 that real data is indispensable if robot models are to generalize in the physical world as broadly as language models do in the virtual oneSource 36. Closed-loop learning from any of these sources still meets the covariate shift covered in behavior cloning foundations.

How to read the field's data claims

Convert everything to hours: AgiBot World's million trajectories are 2,976 hours, less than Ego4D's 3,670 hours of passive videoSource 13. Ask for the breakdown by environment, object, task and operator style; those axes outweigh the hour countSource 9. Treat company-reported scale as a claim until the data or an audit is public: AgiBot World's 30% improvement over OXE pretraining has no independent replicationSource 6.