LIBERO Benchmarks

Cross-source consensus on LIBERO Benchmarks from 1 sources and 3 claims.

1 sources · 3 claims

Uses

The compute breakdown — approximately 78% gradient, 21% rollout — is consistent across all three LIBERO benchmarks tested. — Learn Where Outcomes Diverge: Efficient VLA RL via Probabilistic Chunk Masking
Three LIBERO suites are used for evaluation: LIBERO-Object for object-centric pick-and-place, LIBERO-Spatial for spatial-relation following, and LIBERO-Goal for goal-conditioned manipulation. — Learn Where Outcomes Diverge: Efficient VLA RL via Probabilistic Chunk Masking
The evaluation is limited to single-arm, short-horizon tasks; longer-horizon and bimanual coordination scenarios are not tested. — Learn Where Outcomes Diverge: Efficient VLA RL via Probabilistic Chunk Masking