Glossary
What Is a Pretraining Corpus in Robotics?
Last reviewed: September 2026
Short answer
A pretraining corpus is the large-scale dataset — egocentric human video, robot demonstrations, or a mix of both — that a robot foundation model trains on during pretraining, before it is adapted to any specific task. A corpus is defined by breadth and scale rather than task performance: it aims to cover many scenes, tasks, objects, and people so the model learns general representations that later stages (mid-training and post-training) can specialize.
01
Why it matters
A pretraining corpus’s composition sets a ceiling on what a robot foundation model can later generalize to, because mid-training and post-training only specialize what pretraining already covers. Published corpora vary by roughly two orders of magnitude in scale — EgoVerse reports 1,362 hours, Egocentric-10K reports roughly 10,000 hours, EgoScale reports 20,854 hours, and GEN-0 reports 270,000+ hours — and NVIDIA’s EgoScale paper reports a log-linear scaling law between pretraining data volume and downstream task performance. Generalist’s GEN-0 official blog frames data quality and diversity as mattering as much as sheer hours, so a corpus’s value comes from its coverage and composition, not just its total size.
02
How Firsthand handles it
Firsthand captures the synchronized egocentric video, depth, 3D hand pose, and action-segment data that populate a pretraining corpus, collected fast, on demand, and to a written spec, then delivered in RLDS, LeRobot format, or WebDataset with a buyer-owned commercial license. Because a corpus’s value comes from coverage — scenes, tasks, objects, and people — rather than any single volume target, a custom collection can be scoped around exactly the gaps a public corpus does not fill, rather than adding undifferentiated hours to a layer that is already well represented.
03
Pretraining corpus vs. data mixture vs. fine-tuning set, at a glance
| Term | What it is | How it’s used |
|---|---|---|
| Pretraining corpus | The large-scale dataset a model trains on during pretraining | Consumed broadly, for coverage and general representation learning |
| Data mixture | The blend of sources and their proportions that make up a corpus (e.g. the GR00T N1 pyramid’s layers) | Decided before training, to balance scale against embodiment specificity |
| Fine-tuning (post-training) set | A small, targeted dataset for a specific deployment task | Applied after pretraining, to specialize the already-general model |
05
What typically goes into a pretraining corpus
| Component | What it contributes | Typical source |
|---|---|---|
| Egocentric human video | Broad, easily scaled coverage of scenes, tasks, and hand-object interaction | Human contributors wearing or holding a capture rig, e.g. Firsthand’s stereo capture |
| Synthetic or simulated robot data | Robot-specific dynamics at lower cost than physical capture | Simulation environments and procedurally generated scenes |
| Real robot teleoperation data | Exact actuator and sensor characteristics on the target hardware | A physical robot operated by a human teleoperator |
06
Published pretraining corpora, by scale
| Corpus | Reported scale | Source |
|---|---|---|
| EgoVerse | 1,362 hours | arXiv 2604.07607 |
| Egocentric-10K | ~10,000 hours | Hugging Face (builddotai/Egocentric-10K) |
| EgoScale | 20,854 hours (pretraining) | arXiv 2602.16710 / NVIDIA GEAR |
| GEN-0 | 270,000+ hours | Generalist AI official blog |
07
Frequently asked questions
- What is a pretraining corpus in robotics?
- A pretraining corpus is the large-scale dataset of egocentric human video, robot demonstrations, or both, that a robot foundation model trains on during its pretraining stage, before any task-specific adaptation.
- What goes into a robot pretraining corpus?
- Typically egocentric human video at the base for scale and coverage, plus smaller amounts of synthetic or simulated robot data and real robot teleoperation data — the layering NVIDIA’s GR00T N1 describes as a data pyramid.
- How big are public robot pretraining corpora?
- Reported scales vary widely: EgoVerse reports 1,362 hours, Egocentric-10K reports roughly 10,000 hours, EgoScale reports 20,854 hours, and GEN-0 reports 270,000+ hours. See how much pretraining data does a robot need for the full comparison.
- How is a pretraining corpus different from a data mixture?
- A pretraining corpus is the dataset itself. A data mixture is the blend of sources and their proportions used to build that corpus — for example, how much egocentric human video versus synthetic versus real robot data goes into it.
- How is a pretraining corpus different from a fine-tuning set?
- A pretraining corpus is large and broad, aimed at general coverage across scenes, tasks, and objects. A fine-tuning (post-training) set is small and targeted, scoped to a single deployment task and its edge cases, and applied after pretraining.
- What makes a pretraining corpus useful for robot foundation models?
- Coverage and diversity across scenes, tasks, objects, and people — not just total hours. EgoScale reports a log-linear scaling law between pretraining volume and downstream performance, while GEN-0’s official blog frames data quality and diversity as mattering as much as raw hour count.
- Does a bigger pretraining corpus always produce a better model?
- Not by itself. See do robot foundation models follow data scaling laws for how both EgoScale and GEN-0 frame data composition and diversity as factors alongside raw hours, rather than hours alone.
- How do you build a custom pretraining corpus?
- By scoping a written skill spec around the coverage gaps a public corpus does not fill — specific scenes, tasks, objects, or people — then running a custom collection to that spec, rather than adding undifferentiated volume to an already well-covered layer.
Related terms