0.54M-parameter robot policy hits 95% on LIBERO, 7700x smaller than π0.5

ChongZzZhang · x · 2026-09-09

Key findings

MINERVA is a family of deliberately tiny visuomotor policies built to measure how much model capacity LIBERO actually requires:

A 0.54M-parameter policy hits 95.1% average success over 2,000 rollouts — only 2.4 points below LeRobot π0.5 with 7700x fewer parameters; performance saturates near 1M and collapses below 0.25M.
Across broad architectural/training/inference sweeps, only action-chunk length and vision capacity consistently matter; flow matching shows no advantage over direct L1 regression, which is up to 3.8x faster on GPU.
LIBERO largely tests memorization: permuting only the task-ID mapping drops success to near chance, meaning instruction conditioning mostly selects among memorized tasks.
The same recipe scores 94.6% on 89 LIBERO-90 tasks, but LIBERO-Plus perturbations drop it to 46–56%, with near-zero robustness to photometric shifts.
The 0.54M policy replans every control step in 5–9ms per chunk on a laptop CPU, 113x faster than SmolVLA.

The poster notes that until open-vocab zero-shot setups genuinely advance, VLAs are largely still struggling to align basic representations — high LIBERO scores may not reflect generalizable manipulation.

Original post →

0.54M-parameter robot policy hits 95% on LIBERO, 7700x smaller than π0.5

More from Embodied