## Does Your VLA Policy Actually Know Which Hand to Use?

A study published today on arXiv by Chaeyeon Jung and Juyoun Park identifies a concrete, measurable failure mode lurking inside [Vision-Language-Action Model](https://humanoidintel.ai/glossary/vision-language-action-model) deployments for humanoid dual-arm manipulation: the **policy-induced hand prior**. Evaluating multiple VLA policies across 17 distinct initial arm configurations, the researchers found that aggregate task success rates conceal pose-specific failures severe enough to cause inappropriate hand selection — a robot reaching with the wrong arm not because it misread the scene, but because its starting posture biased the policy's output distribution before any meaningful visual reasoning occurred.

The core finding is direct: a single initial arm configuration can both suppress and induce asymmetric hand preference, and the direction and magnitude of that effect vary across different policies trained on the same task. In other words, the problem is not hardware determinism — it is a learned behavioral artifact baked into the policy during training. The remedy the paper identifies is also concrete: expanding initial-pose coverage in the training dataset substantially improves robustness, while targeted data augmentation around a specific low-performing configuration measurably increases its success rate.

For teams deploying humanoid dual-arm systems in production, this paper is a methodological audit tool as much as it is a research contribution.

---

## What Is a Policy-Induced Hand Prior?

The term the authors introduce — **hand prior** — describes an initial-condition-dependent early preference for one hand over the other that is not driven by task semantics but by the robot's starting configuration. They quantify this phenomenon using three metrics defined in the paper: **HandPriorScore**, **residual hand bias**, and **target responsiveness**.

HandPriorScore captures whether the policy's hand selection correlates with initial pose rather than with the task object's location or affordance. Residual hand bias measures the asymmetry in hand usage after controlling for task-relevant factors. Target responsiveness assesses how strongly the policy adjusts hand selection in response to where the manipulation target actually is.

Together, these three metrics form a diagnostic framework that aggregate success rate — the number most teams report — entirely fails to surface. A policy can clear a 70%+ success rate on a dual-arm pick-and-place benchmark while exhibiting a strong hand prior on specific initial configurations that would be catastrophic in deployment variance. This is not a hypothetical: the paper's cross-policy, cross-pose evaluation directly demonstrates it.

The analogy for engineers familiar with [sim-to-real transfer](https://humanoidintel.ai/glossary/sim-to-real-transfer) is instructive. Just as domain randomization in simulation can paper over actuator-level dynamics mismatches that only surface under real-world load, reporting only aggregate task success can paper over configuration-level policy failures that only surface when a robot arm starts from an underrepresented pose.

---

## The 17-Configuration Experiment and What It Reveals

The study's evaluation backbone is a sweep across 17 initial arm configurations, testing multiple VLA policies on each. The findings have two distinct layers worth separating:

**Inter-policy variation at the same pose:** The same starting configuration produces substantially different success rates across policies. This tells you that the hand prior is not purely a function of kinematics — it is a function of what each policy learned from its training data distribution. Two policies trained to accomplish the same task can develop opposite hand preferences when initialized identically.

**Intra-policy variation across poses:** A single policy exhibits large performance swings across the 17 initial configurations. This is the operationally dangerous finding. A policy validated at a standard initial pose may degrade sharply when deployed on hardware that settles into a slightly different resting configuration between tasks, or when a human repositions a robot arm before a handoff.

The paper also identifies **wrist-camera observations** as an additional confound: the perspective from the wrist camera influences both hand selection and overall task performance, adding a sensor-geometry dimension to the pose-dependence problem.

Critically, the authors do not stop at diagnosis. They test interventions:

- **Expanding initial-pose coverage** in the training dataset substantially improves cross-configuration robustness.
- **Targeted augmentation** around a specifically low-performing configuration increases that configuration's success rate.
- **Training composition effects** are nuanced: sufficient exposure to the target simulation task is beneficial, but the effect of real-world or auxiliary data depends on pose coverage, simulation ratio, and observation modality availability.

That last point deserves emphasis for any team currently mixing sim and real data. More real data is not unconditionally better — if that real data was collected from a narrow slice of initial poses, it can reinforce rather than correct a hand prior.

---

## Why This Matters for Humanoid Deployment at Scale

[Dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation) in humanoid dual-arm systems is already one of the hardest open problems in the field. The engineering community has focused heavily on hand hardware — actuator backdrivability, finger DOF counts, tactile sensing — and increasingly on large-scale [imitation learning](https://humanoidintel.ai/glossary/imitation-learning) pipelines. What this paper surfaces is a data curation and evaluation methodology gap that sits between those two domains.

The practical implication: **benchmark protocols for VLA-based manipulation need to explicitly include initial-pose variation as a first-class variable, not an afterthought.** Reporting a single success rate from a canonical starting pose is insufficient for any claim of deployment readiness. The HandPriorScore, residual hand bias, and target responsiveness metrics introduced here give the field concrete tools to do better.

For companies like [Physical Intelligence (π)](https://humanoidintel.ai/companies/physical-intelligence) and [Skild AI](https://humanoidintel.ai/companies/skild-ai) building generalist manipulation policies intended to run on multiple humanoid platforms, the pose-coverage question is particularly acute. A policy that generalizes across robot embodiments but not across initial configurations of a single embodiment has a fundamental robustness gap — one that would not appear in standard [zero-shot generalization](https://humanoidintel.ai/glossary/zero-shot-generalization) evaluations that fix the starting pose.

The skeptical read: the paper evaluates "multiple policies" without naming them or providing hardware platform specifics in the abstract. The 17-configuration sweep is a meaningful sample but not exhaustive. Whether HandPriorScore generalizes cleanly to contact-rich tasks — assembly, tool use, bimanual cloth manipulation — beyond the evaluated scenarios remains an open question. But the diagnostic framework itself is sound, and the training-data intervention results are directly actionable.

---

## Key Takeaways

- **Hand priors are policy artifacts, not hardware artifacts.** The same initial arm configuration produces different hand-selection behavior across different VLA policies, confirming the bias is learned, not kinematically determined.
- **Aggregate success rates hide pose-specific failures.** A policy can appear competent on standard benchmarks while catastrophically misbehaving on specific initial configurations.
- **Three new diagnostic metrics** — HandPriorScore, residual hand bias, and target responsiveness — give teams concrete tools to audit their policies beyond success rate.
- **Wrist-camera geometry is a confound.** Camera perspective from the wrist influences hand selection and task performance independently of arm kinematics.
- **Data fixes work, but require pose-aware curation.** Expanding initial-pose coverage improves robustness; more real data without pose diversity does not.
- **Benchmark protocols need updating.** Initial-pose variation should be a required axis of evaluation for any dual-arm humanoid manipulation claim.

---

## Frequently Asked Questions

**What is a policy-induced hand prior in robotics?**
A policy-induced hand prior is a learned bias in a VLA policy that causes it to prefer one hand over the other based on the robot's initial arm configuration rather than the task's actual requirements. Jung and Park (2026) define and quantify this using HandPriorScore, residual hand bias, and target responsiveness metrics evaluated across 17 initial configurations.

**How does initial pose affect VLA policy performance in dual-arm manipulation?**
According to the arXiv paper (2608.11769), a single initial arm configuration can substantially change success rates for the same policy, and the same pose can produce very different outcomes across different policies. Specific configurations can suppress or induce asymmetric hand preference, creating performance variation that aggregate success metrics hide entirely.

**Does adding more real-world training data fix the hand prior problem?**
Not unconditionally. The paper finds that the effect of real or auxiliary data on initial-pose robustness depends on pose coverage, simulation ratio, and observation availability. Real data collected from a narrow range of starting poses can reinforce rather than correct a hand prior.

**What is HandPriorScore and how is it used?**
HandPriorScore is a metric introduced by Jung and Park to quantify how strongly a VLA policy's hand selection correlates with the robot's initial configuration rather than with task-relevant factors. It is used alongside residual hand bias and target responsiveness to diagnose pose-dependent hand selection failures that success rate alone cannot reveal.

**How should humanoid robotics teams change their evaluation benchmarks based on this research?**
Teams should treat initial arm configuration as a first-class evaluation variable rather than fixing a canonical starting pose. Running policy evaluations across a diverse set of initial configurations — and reporting HandPriorScore alongside task success rate — would surface deployment risks that current benchmark protocols systematically miss.