robotics
Five Percent Exposure Changed What Zero-Shot Meant
ZETA separates truly unseen robot hardware from embodiments glimpsed during pretraining across 14 held-out targets.
Summary
ZETA separates truly unseen robot hardware from embodiments glimpsed during pretraining across 14 held-out targets.
The benchmark holds tasks, scenes and protocols steady while changing robot embodiment, distinguishing strict zero-shot transfer from pretrain-exposed transfer. Local end-effector representations, source-embodiment diversity and auxiliary co-training improved average transfer by about 15, 18 and 7 percentage points in the reported analysis. Adding only 5% target-embodiment data during pretraining raised target progress by 13.4 points, showing why the two zero-shot conditions should not share one label.
Why it matters
ZETA separates truly unseen robot hardware from embodiments glimpsed during pretraining across 14 held-out targets.
Limits and context
- Adding only 5% target-embodiment data during pretraining raised target progress by 13.4 points, showing why the two zero-shot conditions should not share one label.
Key claims
ZETA separates truly unseen robot hardware from embodiments glimpsed during pretraining across 14 held-out targets.
Qualification: Adding only 5% target-embodiment data during pretraining raised target progress by 13.4 points, showing why the two zero-shot conditions should not share one label.
Evidence: source-2026-09-03-012
Sources
- arXiv preprint 2609.02546arXiv · primary research
Corrections
No corrections have been recorded for this story.