# Does Wrapping a VLA in an Agentic Loop Solve Generalist Robot Failures?

A new paper from Liane Galanti, Dhruv Shah, and Tri Dao answers that question with a single striking number: a frozen [vision-language-action model](https://humanoidintel.ai/glossary/vision-language-action-model) policy, unchanged and unretrained, jumps from 12.8% to 53.3% task success on the LIBERO-PRO benchmark — a greater than 4x improvement — simply by wrapping it inside their Physical Agency orchestrator, which the authors call **Pigey**. On a real robot, Pigey lifts the same frozen policy from near-zero success to over 90% on what they classify as reasoning-limited tasks, again with no additional data collection or post-training.

The paper, published today on arXiv (2607.21725), introduces the concept of the **orchestration gap**: the measurable delta between what frozen motor skills achieve when run alone versus when those same skills are embedded inside a closed-loop agentic system. That framing matters. It reframes a large class of perceived policy failures not as a learning problem, but as an architecture problem — and that distinction has direct implications for every lab currently burning GPU hours on large-scale pre-training in hopes of coaxing more reasoning out of learned policies.

---

## What Is the Orchestration Gap, Exactly?

The standard paradigm in generalist robot [physical AI](https://humanoidintel.ai/glossary/physical-ai) has been to train a single, monolithic model end-to-end to handle perception, planning, task decomposition, motor execution, and failure recovery simultaneously. The bet is that scale — more data, more parameters, more compute — will eventually unify these capabilities. Pigey's authors argue this conflates capabilities that are architecturally separable, and that the conflation is directly responsible for the brittleness practitioners observe in deployed VLA systems.

Their decomposition is straightforward:

- **Low-level control agent**: a general language-conditioned policy (an existing VLA or parameterized skill set), left entirely frozen.
- **High-level orchestrator (Pigey)**: responsible for goal decomposition into achievable subgoals, issuing motor commands, monitoring execution from low-level observations, verifying success, and triggering recovery when tasks stall or fail.

Critically, Pigey requires no task-specific fine-tuning. It is designed to wrap *existing* policies — a direct challenge to the assumption that better task performance requires better (i.e., more expensively trained) policies.

---

## The LIBERO-PRO Numbers in Context

The LIBERO-PRO benchmark is a meaningful stress test because it is specifically designed to probe reasoning and long-horizon task completion, not just atomic manipulation primitives. The reported improvement — from 12.8% to 53.3% — is not a marginal gain; it represents the system moving from essentially failing at the benchmark to achieving majority success, purely through architectural change.

The real-robot results are arguably more striking. A policy operating at near-zero success on reasoning-limited tasks clearing 90% inside the Pigey loop is the kind of delta that, in a deployment context, is the difference between a system being shelved and one being productized.

The authors evaluated Pigey across both simulation benchmarks and challenging real-world robotic manipulation tasks, providing some sim-to-real validation of the approach, though the paper does not claim [zero-shot generalization](https://humanoidintel.ai/glossary/zero-shot-generalization) across arbitrary environments.

---

## Why This Matters for the Humanoid Software Stack

The humanoid industry's current software investment thesis is heavily weighted toward foundation model scaling — the assumption, broadly shared across labs backed by significant capital, that a sufficiently large VLA trained on sufficiently diverse manipulation data will generalize to novel tasks with acceptable reliability. Companies like [Physical Intelligence (π)](https://humanoidintel.ai/companies/physical-intelligence) and [Skild AI](https://humanoidintel.ai/companies/skild-ai) have raised substantial funding on variants of this thesis.

Pigey's results do not invalidate that thesis, but they complicate it. If a significant fraction of VLA failures are orchestration failures — tasks where the underlying motor skill is adequate but the system cannot decompose goals, detect failure states, or recover — then adding a high-level orchestrator may yield faster near-term capability gains than further policy scaling. This is a resource allocation argument as much as a technical one: orchestration layers are far cheaper to develop and iterate on than pre-training runs.

There is a legitimate counterargument: Pigey's gains are demonstrated on **reasoning-limited tasks**, a category the authors themselves define. If a task fails because the low-level policy genuinely lacks the motor capability, no orchestration layer rescues it. The 90% real-robot result is impressive, but the selection of tasks matters enormously. Independent replication on a broader task distribution will be the real test.

The authors — Galanti, Shah, and Dao — are credible names. Tri Dao in particular is widely known in the ML community for work on efficient attention mechanisms, making his presence on a robotics orchestration paper notable and worth tracking for follow-on work.

---

## Industry Trajectory Implications

The orchestration gap concept, if it holds up under wider scrutiny, reframes how humanoid hardware companies should think about their software procurement. Buying or training better VLA backbones is one lever; deploying an orchestration layer on top of an existing policy is another, potentially cheaper one. For companies at the deployment stage — running pilots in warehouses, logistics facilities, or manufacturing lines — a system that can recover from failures without human intervention is often the difference between a viable product and a supervised demo.

The closed-loop verification component is particularly relevant for [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation) use cases. The ability to detect when a grasp has failed and re-attempt without human reset is a known pain point in real deployments, and Pigey's architecture directly addresses it at the system level rather than requiring the policy to learn it implicitly.

---

## Key Takeaways

- **Pigey** is a closed-loop physical agent orchestrator wrapping existing VLA policies, requiring no additional data collection or post-training.
- On LIBERO-PRO, it improves state-of-the-art task success from **12.8% to 53.3%** — over 4x — with no task-specific fine-tuning.
- On a real robot, it lifts a frozen policy from **near-zero to over 90%** on reasoning-limited tasks.
- The paper introduces the term **orchestration gap** to describe the delta between what frozen motor skills achieve alone versus inside an agentic loop.
- The approach challenges the dominant assumption that better generalist performance requires better (more expensive) pre-trained policies.
- The framework works with both VLA policies and parameterized skill sets, suggesting broad compatibility with existing stacks.
- Independent validation on diverse, non-curated task distributions remains the critical open question.

---

## Frequently Asked Questions

**What is the orchestration gap in robotics?**
The orchestration gap, as defined by Galanti, Shah, and Dao in arXiv:2607.21725, is the measurable difference in task success between a frozen motor skill policy operating alone and the same policy operating inside a closed-loop agentic orchestration system. The gap represents failures attributable to missing high-level reasoning, planning, and recovery capabilities rather than to deficiencies in the underlying motor policy itself.

**What is Pigey and how does it work?**
Pigey (Physical Agency orchestrator) is a high-level agent manager that wraps existing vision-language-action policies or parameterized skills without modifying or retraining them. It handles goal decomposition into subgoals, issues commands to the low-level policy, monitors execution outcomes from sensor observations, verifies task success, and initiates recovery when failures are detected — all in a closed loop.

**What benchmark results did Pigey achieve?**
On the LIBERO-PRO simulation benchmark, Pigey advanced state-of-the-art performance from 12.8% to 53.3% task success with no task-specific fine-tuning. On a real robot, it lifted a frozen policy from near-zero to over 90% success on reasoning-limited tasks.

**Does Pigey require retraining the underlying policy?**
No. A central claim of the paper is that Pigey achieves its performance gains without any additional data collection or post-training of the underlying policy. The low-level control policy remains frozen throughout.

**How does this affect the case for scaling VLA pre-training?**
Pigey's results suggest that a meaningful fraction of VLA task failures are orchestration failures rather than motor capability failures. This implies that adding an orchestration layer may be a more cost-efficient near-term path to improved deployment reliability than further scaling pre-training runs — though the two approaches are not mutually exclusive, and the extent of the gains on broader, non-curated task distributions remains to be established.