"Today's robot foundation models run to billions of parameters and stacks of top-tier GPUs. So flip the question: how much compute does a robot minimally need to get the job done?"
MINERVA, from the Matsuo Lab at the University of Tokyo (arXiv 2609.03715, posted September 3), doesn't follow the "how do we shrink big models" path. Instead it measures in reverse: taking a manipulation policy from 9.7M all the way down to 0.24M parameters, retraining and re-evaluating after each cut, watching where the task suite breaks down. The last surviving point before the break is the capacity floor.
The answer: 0.54M parameters. Across the four LIBERO suites, 2000 evaluations, average success rate 95.1% — only 2.4 points below the 4.1-billion-parameter pi0.5 (per the official LeRobot implementation). A 7700x parameter gap.
First, verify the numbers
We checked every figure in the source material against the paper itself: 0.54M ✓, 95.1% (2000 rollouts) ✓, 7700x ✓, the 2.4-point gap ✓, SmolVLA 1.0 second per chunk ✓, pi0.5 12.8 seconds ✓, 5–9 ms per chunk on a laptop CPU ✓.
Two caveats on framing: first, the 1.0s vs 12.8s comparison is on the same laptop CPU (paper Table III) — the source omitted that context. pi0.5 is not slow on GPU hardware; we previously recorded it hitting 24.3Hz on Jetson Thor FP8. What this comparison really says is "can you work without a GPU." Second, calling it a "hot new study" needs a discount: one month after posting, the GitHub repo has 19 stars and zero media coverage. Not a hit — a should-be-hit that hasn't broken out yet. No one in the Chinese-language community had covered it; this piece is a first.
It's not a smaller VLA — it's a VLA taken apart
The 7700x gap wasn't achieved through compression. Of pi0.5's three pillars — pretrained vision-language backbone, language encoder, self-attention — MINERVA keeps none:
Vision is a CNN trained from scratch, attached to no pretrained backbone. Language is deleted outright, replaced by a learned ID badge per task (40 task-ID embeddings) — LIBERO's standard suite has 40 tasks; you don't need to understand language, just recognize the badge. Attention becomes an MLP mixer, distilled from a teacher model.
Only two moving parts survive: a flow-matching action-chunk head, and a tunable visual budget. The paper sweeps hundreds of architecture, training, and inference combinations; only two factors stably exceed training-seed variance (±1 point): action chunk length (8 is too short, −3.2 points; 32 too long, −2.2) and visual budget (starving the encoder costs −1.41 points). Everything else is noise.
Three cuts the paper makes against itself
This is where the paper is worth more than its numbers — it discloses the quality of its own report card.
Cut one: perturbations halve the score. On the perturbed LIBEROPlus evaluation, success drops to 46–56%, and photometric robustness (lighting, color shifts) is near zero at every scale — not a small-model disease; the 4B giants have it too. 95.1% is the "exam as originally printed" score.
Cut two: instruction conditioning is mostly memorization. The researchers ran a permutation probe: shuffle only the task-ID mapping, touch nothing else, and success collapses to random level. The conclusion is in the abstract: on LIBERO, language instructions serve to select among 40 memorized tasks. The instructions didn't teach the robot language — they just help it look up its exam number.
Cut three: flow matching has no advantage. The signature technique of the pi0 series shows no measurable gain over three seeds on a fixed task set; plain L1 regression matches the score and is 3.8x faster on GPU.
The capacity floor itself is double-bounded: above 1M parameters performance saturates (bigger is waste), below 0.25M it collapses (of the 0.24M row, the paper's word is *breaks*).
Fitting the main storyline
The deployment-collapse spectrum gains another pole. Prior tiers all "compressed existing models" (quantization, distillation, speculative decoding). MINERVA changes the question: not "how small can we compress a model" but "how much capacity does a fixed task set actually need." Compression spends carefully; the capacity floor designs to need. On the roadmap, it and TurboVLA (0.2B, 97.7%, 32Hz on RTX 4090 — we covered it in August) are two rungs of the same direction: TurboVLA still keeps a language pathway; MINERVA stripped out language entirely.
The memorization probe carries broader meaning too. The duty to distinguish "memorizing vs solving" was fulfilled by the paper itself this time. LIBERO score inflation is an industry phenomenon (everyone sits at saturated 95+), and this paper punctures the ruler in passing: another reading of score saturation is that the ruler rewards memorization. A falsifiable prediction: within twelve months, standard evaluation tables in manipulation papers will add a perturbation-results column, and LIBERO scores without LIBEROPlus will be filed separately as "original-exam" scores.
End on the 0.24M row. Just 2x from the floor, all tasks collapse. The other side of the floor is not a slope but a cliff — which also explains why pi0.5 needs 4.1 billion parameters: it isn't guarding against these 40 LIBERO questions; it's guarding against every world that hasn't issued an exam paper yet.
---
Sources: the paper itself (arXiv 2609.03715, checked section by section; 2000 rollouts / Table III / permutation probe / LIBEROPlus figures verified in body and abstract), GitHub repo API (19 stars, captured 2026-10-06), alphaxiv cross-check, TurboVLA and pi0.5 Jetson figures from our own 08-19/09-16 posts. The source material's "viral hit" claim was ruled premature based on actual repo counts.