GR00T N1.6 fine-tuned twice, once on SABER-10K as released and once on an edited version with the lowest-scoring quarter removed. Same budget, same ten retail tasks, 100 rollouts each. The edited version won.
A same-size random cut scored 30.1%, so the gain is from which episodes were removed, not from removing episodes.
The tasks that gained most are the ones the full corpus had left flat. The pooled lift clears zero; most single tasks at 100 rollouts do not, and the chart says so.
Every episode was scored by how well its own frames predict its own actions, against a shuffled copy of itself. The lowest quarter was dropped. It is not a random quarter. 2,140 of the 2,500 removed episodes come from the egocentric stream, and in 1,900 of those the latent action is near constant for more than 80% of the clip while the camera keeps moving. They are aisle-walking segments labelled as manipulation. There is no action in them for a policy to learn, and the score found them without being told.
| Removed quarter | episodes | margin |
|---|---|---|
| aisle walking, near-constant latent action | 1,900 | −4% |
| hand out of frame for most of the clip | 312 | −1% |
| retarget failure, joints pinned at limits | 204 | −6% |
| other low-margin | 84 | +2% |
| Removed | 2,500 | −3% |
| Kept | 7,500 | +14% |
Measured on the release. The egocentric stream's action is a four-token code, each token 0 to 7. 565 of its 5,000 episodes hold the same code on 80% or more of their frames, and the median episode repeats it on 71%. Seven in ten of its task strings describe walking, turning or pushing a cart. These six are the most constant in the whole stream. The last clip is one the cut keeps, for contrast.