Guide
How much egocentric data do you need to train a robot policy?
Short answer
As a starting point, budget 50–100 validated hours per skill, concentrated on the conditions your policy will actually meet. Returns are roughly log-linear, so the first hundred hours move the needle most. Scale further only where held-out evaluation shows the policy is still thin.
Hours of what, exactly?
The unit is validated hours, not raw recording. Teams routinely discard most bulk egocentric footage as unfit for manipulation, so a raw-hour budget overstates what you actually get to train on. Budget and evaluate in validated hours — footage that passed QA against your spec — and the number becomes meaningful.
How do returns scale with hours?
Empirically, task-completion gains are roughly log-linear in demonstration hours: the jump from 10 to 100 hours dwarfs the jump from 1,000 to 1,090. That is why a tightly-scoped first 50–100 hours per skill is the highest-leverage spend, and why coverage of the hard conditions beats raw volume.
The EgoScale scaling study reports the same log-linear relationship; use it as an external reference point.
A concrete budgeting approach
- Pick one skill and write its spec, including condition cells.
- Collect ~50 validated hours weighted toward failure cases.
- Evaluate on held-out environments; find the thin cells.
- Add hours only into the cells evaluation says are weak.
Check it against the sample pack.
40 episodes across 4 environments, delivered in the exact schema these guides describe.