# Does a Humanoid Manipulation Policy Need Billions of Parameters?

**0.54 million parameters.** That is all MINERVA needs to achieve 95.1% average success across the four standard LIBERO suites — a result that sits only 2.4 percentage points below the reported [Physical Intelligence (π)](https://humanoidintel.ai/companies/physical-intelligence) LeRobot π0.5 score, despite using 7,700× fewer parameters. The paper, published today on arXiv by Kohei Sendai, Tatsuya Matsushima, and Yusuke Iwasawa, delivers one of the most pointed capacity audits the manipulation policy literature has seen.

The implication is uncomfortable for the VLA field: the benchmark that billion-parameter [vision-language-action models](https://humanoidintel.ai/glossary/vision-language-action-model) are winning may not require anywhere near that capacity to beat. MINERVA establishes what the authors call a "task-specific capacity floor" — and that floor turns out to be surprisingly low.

For engineers building deployment-constrained humanoid systems, the headline inference number matters just as much as the accuracy figure: the 0.54M policy replans every control step in 5–9 ms per chunk on a **laptop CPU**, without a GPU. That is 113× faster than SmolVLA and 1,400× faster than π0.5, per the paper's reported comparisons.

---

## What MINERVA Actually Measures

LIBERO is the standard tabletop manipulation benchmark suite used widely to compare visuomotor policies. The benchmark comprises four task suites, and the authors evaluated MINERVA over 2,000 rollouts — a sample size large enough to produce statistically meaningful results.

The MINERVA family is described as "deliberately compact," built specifically to probe the task-specific capacity floor rather than to chase leaderboard position with scale. Performance, per the paper, saturates near 1M parameters and **collapses below 0.25M** — giving researchers a reasonably tight bracket: roughly 0.25M–1M parameters is all LIBERO appears to demand.

The same recipe applied to LIBERO-90 achieved 94.6% success across 89 tasks — a consistent result that reinforces the core finding.

---

## The Architectural Findings That Should Shift Design Decisions

Three findings from the paper's broad architectural sweep deserve particular attention from practitioners:

**1. Flow matching offers no detectable advantage over L1 regression on LIBERO.** Across three seeds, the paper reports no performance edge for flow matching — while direct L1 regression is up to 3.8× faster on GPU. This is a direct challenge to a design assumption that has proliferated across recent policy architectures. Flow matching is computationally heavier and, at least on this benchmark, buys nothing in success rate.

**2. Only two hyperparameters consistently exceed a ±1-point training-seed band: action-chunk length and vision capacity.** Across the paper's "broad architectural, training, and inference sweeps," everything else — architecture variants, training choices, inference settings — stays within the noise. This is a meaningful prior for teams spending engineering time on policy tuning.

**3. LIBERO's language conditioning is largely task ID lookup, not semantic generalization.** The most pointed finding may be the task-ID permutation probe. When the authors changed only the task-ID mapping, success dropped to near chance. The conclusion: standard LIBERO instruction conditioning primarily selects among memorized tasks rather than demonstrating genuine language understanding or [zero-shot generalization](https://humanoidintel.ai/glossary/zero-shot-generalization). The benchmark is, to a significant degree, a memorization test.

---

## The Robustness Gap Is the Real Deployment Problem

Here is where the MINERVA results become sobering for anyone considering LIBERO performance as a proxy for real-world readiness.

LIBERO-Plus perturbations — described in the paper as environmental variations — reduce MINERVA's performance to 46–56%. More critically: the policy shows **near-zero robustness to photometric shifts**. Change the lighting or image appearance, and success collapses.

This is not a MINERVA-specific weakness. It is a LIBERO-specific weakness that MINERVA has now made explicit. If a 0.54M parameter policy can match a billion-parameter VLA on this benchmark, and both likely share the same fragility to photometric perturbation, then LIBERO success rates tell us very little about what either policy would do on a real humanoid in an uncontrolled environment.

For robotics engineers designing [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation) stacks for humanoid deployment, this is the number to internalize: 46–56% under perturbation, near-zero under photometric shift. That is the actual performance envelope LIBERO measures — and it is far narrower than leaderboard scores suggest.

---

## Industry Trajectory: Capacity-Aware Design and Distillation

The MINERVA paper is not arguing that billion-parameter VLAs are useless — they likely earn their parameter counts on harder, more semantically diverse tasks. The argument is more specific: **benchmark-driven scale is not the same as deployment-necessary scale**, and the field has conflated the two.

The practical upshot for humanoid AI teams: if your manipulation task distribution is bounded and your training data covers it well, a sub-million-parameter policy running on CPU may outperform a cloud-dependent VLA on latency, cost, and on-device feasibility. The 5–9 ms per chunk replanning figure is particularly relevant for whole-body control integration, where policy latency directly affects stability.

The authors explicitly motivate "capacity-aware design and distillation for deployment-efficient robot policies" — which points toward a workflow where large VLAs are used to generate demonstrations or teacher signals, then distilled into compact deployment policies. That pipeline already exists in spirit across several humanoid programs, but MINERVA provides empirical grounding for how aggressively you can compress without losing task performance.

The deeper question the paper raises — whether LIBERO should remain the field's primary manipulation benchmark given its memorization dynamics and photometric fragility — is one the community will need to answer with harder benchmarks and real-hardware evaluation.

---

## Key Takeaways

- **MINERVA's 0.54M-parameter policy achieves 95.1% on LIBERO's four standard suites**, only 2.4 points below π0.5 with 7,700× fewer parameters
- **Performance saturates near 1M parameters and collapses below 0.25M** — establishing an empirical capacity floor for LIBERO
- **Flow matching provides no detectable advantage over L1 regression** on this benchmark; regression is up to 3.8× faster on GPU
- **Only action-chunk length and vision capacity consistently matter** across broad architectural sweeps
- **LIBERO instruction conditioning primarily selects memorized tasks** — a permutation probe drops success to near chance
- **LIBERO-Plus perturbations reduce performance to 46–56%**; photometric robustness is near zero — a benchmark validity problem, not just a model problem
- **The 0.54M policy runs at 5–9 ms per chunk on a laptop CPU**, 1,400× faster than π0.5
- **Implication for deployment:** capacity-aware distillation from large VLAs into compact policies is empirically motivated

---

## Frequently Asked Questions

**What is MINERVA in robotics?**
MINERVA (MINimal Efficient Robotic Vision-Action policy) is a family of compact visuomotor policies introduced by Sendai, Matsushima, and Iwasawa to measure the minimum model capacity needed to solve the LIBERO manipulation benchmark. Its 0.54M-parameter variant achieves 95.1% average success across LIBERO's four standard suites.

**How does MINERVA compare to π0.5 on LIBERO?**
The paper reports MINERVA's 0.54M policy scores 2.4 percentage points below the reported LeRobot π0.5 result, while using 7,700× fewer parameters and running 1,400× faster on inference — without requiring a GPU.

**Does LIBERO actually test generalization?**
The MINERVA paper's task-ID permutation probe suggests standard LIBERO instruction conditioning primarily selects among memorized tasks rather than generalizing from language semantics. Changing only the task-ID mapping drops success to near chance, indicating the benchmark has significant memorization dynamics.

**Is flow matching better than L1 regression for manipulation policies?**
According to MINERVA's experiments across three seeds, flow matching provides no detectable success-rate advantage over direct L1 regression on LIBERO, while L1 regression is up to 3.8× faster on GPU.

**What is LIBERO-Plus and why does it matter?**
LIBERO-Plus introduces perturbations to the standard LIBERO environment. MINERVA's performance drops to 46–56% under these perturbations, with near-zero robustness to photometric shifts — highlighting that standard LIBERO scores may substantially overstate real-world manipulation capability.