BruinML · Research · Fall 2026–present
Rollouts Needed to Rank
How much evaluation is enough to tell which demonstrations are most useful to a learned robot policy?
- Status
- Ongoing · started Fall 2026
- Area
- Learning theory & robotics
- Focus
- Demonstration influence & ranking
The question
Demonstrations provide the examples a robot learns from. Rollouts are new attempts made by the trained policy. This project asks how many of those evaluation attempts are needed to reliably rank demonstrations by their influence on expected task performance.
Why ranking is difficult
Influence scores estimated from a small rollout budget are noisy. When demonstrations have similar scores, that noise can change which ones appear most useful. Reusing rollouts also makes the estimates statistically dependent, so confidence in a ranking depends on more than each score in isolation.
Research direction
The project will study how rollout budget, estimator variance, and gaps between influence scores affect reliable selection. Variance-reduction methods are one direction of interest. A second distinction matters throughout: estimating an influence ranking accurately is not the same as proving which subset performs best after retraining.
This entry describes the goal of a newly started project. Individual contributions and results will be added as the work develops.
