X4Val: Variance-Reduced Evaluation

• X4Val is a novel framework for variance-reduced evaluation in robotics, integrating neural surrogates with control variates to combine disparate data sources.
• It leverages simulation, historical policies, and scenario-only target data to enhance sample efficiency, achieving up to 38.4% variance reduction over Monte Carlo.
• Its cross-fitting estimator preserves unbiasedness and provides rigorous confidence intervals, enabling effective transfer across different domains.

Searching arXiv for the X4Val paper and closely related work on control variates / prediction-powered evaluation in robotics. X4Val is a framework for variance-reduced evaluation of robotic and autonomous systems in settings where real-world metric labels are expensive, while heterogeneous auxiliary data are abundant but non-representative of deployment outcomes. It is designed for non-paired, multi-domain data, and combines a learned neural surrogate of the deployment-domain metric with a control-variates estimator that preserves unbiasedness and supports rigorous confidence intervals. The framework is presented for policy evaluation in autonomous driving and robot manipulation, where it uses simulation, historical policies, other platforms, and scenario-only target-domain data to improve sample efficiency; the reported empirical gains reach up to 38.4% variance reduction (Luo et al., 3 Jun 2026).

1. Problem formulation and target quantity

X4Val studies the estimation of a deployment-domain performance metric for a policy $\pi$ over scenarios $X$ drawn from a deployment distribution $P_X$. The target is the mean metric

$\mu \coloneqq \mathbb{E}_{X\sim P_X}[F(X)].$

Here, $X$ denotes a scenario, such as an autonomous-driving clip or a robot initial configuration, and $F(X)$ is a scalar performance metric, such as a success indicator, final displacement error, or distance to a ground-truth trajectory (Luo et al., 3 Jun 2026).

A direct Monte Carlo estimator uses labeled deployment-domain trials,

$$\hat{\mu}_{\text{MC}} = \frac{1}{n}\sum_{i=1}^n F_i,$$

with variance

$$\mathrm{Var}(\hat{\mu}_{\text{MC}}) = \frac{1}{n}\,\mathrm{Var}(F).$$

The statistical bottleneck is that each $F_i$ may require a real-world experiment, a closed-loop simulation, or another expensive evaluation procedure. In iterative robotic development, this is compounded by the fact that data from the latest policy are necessarily limited in scale, even when large historical or simulated datasets are available (Luo et al., 3 Jun 2026).

The framework is motivated by the ubiquity of auxiliary data that are cheap but distributionally shifted relative to deployment: simulation rollouts, logs from earlier policies, evaluations on other platforms or geographies, and target-domain scenarios without expensive labels. The core difficulty is to exploit these sources for variance reduction without introducing bias into the estimate of $\mu$ (Luo et al., 3 Jun 2026).

2. Estimator design: neural surrogate and control variates

X4Val uses three classes of data. The first is labeled target-domain metric data,

$X$0

where $X$1 is cheaply available side information. The second is target-domain scenario-only data,

$X$2

for which the expensive metric is not observed. The third is a collection of auxiliary datasets $X$3 from related but shifted domains (Luo et al., 3 Jun 2026).

All domains are embedded into a shared representation

$X$4

and a surrogate

$X$5

is learned to predict the deployment-domain metric. The single-split X4Val estimator is

$X$6

Its first term estimates the mean surrogate prediction over many cheap target-domain scenarios; its second term estimates the residual correction on labeled target-domain data (Luo et al., 3 Jun 2026).

The critical property is unbiasedness. For any fixed surrogate trained without using labels in $X$7,

$X$8

This means that the surrogate may itself be biased, or trained on shifted auxiliary domains, without biasing the final estimator. What matters is that the residual correction is computed on deployment-domain samples and that label leakage is avoided (Luo et al., 3 Jun 2026).

Conditioning on the surrogate, the estimator variance decomposes as

$X$9

This decomposition motivates X4Val’s emphasis on residual-variance reduction rather than surrogate point prediction alone. The paper therefore proposes a variance-aligned training loss,

$P_X$0

which approximates direct minimization of the estimator variance. A simpler alternative is to introduce a scalar control weight $P_X$1 through $P_X$2 and tune $P_X$3 on held-out target data (Luo et al., 3 Jun 2026).

3. Shared representation learning and transfer across domains

A defining feature of X4Val is that it does not require paired real and auxiliary measurements on the same scenario. Auxiliary domains are used only to learn a transferable surrogate in a shared feature space, after which the final estimator uses target-domain scenario-only samples and target-domain labeled samples (Luo et al., 3 Jun 2026).

The shared representation takes the form $P_X$4, where $P_X$5 may include hand-designed features, learned visual embeddings, simulator quantities, open-loop metrics, or other cheap signals, provided that these are computable in both target and auxiliary domains. In the autonomous-driving experiments, $P_X$6 includes a DINOv3 embedding of the front camera image, ego-vehicle body-frame velocity and acceleration, and sometimes open-loop metrics such as final displacement error. In the robot-manipulation experiments, $P_X$7 is derived from DINOv2 embeddings of initial camera images, yielding a 384-dimensional vector (Luo et al., 3 Jun 2026).

The surrogate architecture is application-dependent. For autonomous driving, the paper uses a 2-layer MLP with ReLU on scenario feature vectors of dimension 772–773. For policy-family transfer, it introduces an amortized meta-learning construction based on a DeepSets encoder. In that setup, a set of scenario-metric pairs associated with a policy is mapped through a per-element MLP, pooled by sum or mean, and processed by an output MLP into a 64-dimensional policy embedding; a shared-backbone predictor is then conditioned on that embedding to predict the target metric. For robot manipulation, the surrogate is a 2-hidden-layer MLP with hidden size 128 on top of fixed DINOv2 features, predicting binary success (Luo et al., 3 Jun 2026).

The training algorithm is written generically as

$P_X$8

and may be instantiated as pretrain-finetune transfer learning, amortized meta-learning across policies, or cross-prediction inside the X4Val estimator. The practical consequence is that non-paired, multi-domain data are admissible even when exact sim-real or policy-policy correspondences are unavailable. This is a substantive departure from classical control-variates approaches that require paired real and simulated measurements on the same scenarios (Luo et al., 3 Jun 2026).

4. Cross-fitting, confidence intervals, and statistical assumptions

To use all labeled target-domain samples while preserving unbiasedness, X4Val employs cross-fitting. The labeled dataset is partitioned into $P_X$9 disjoint folds $\mu \coloneqq \mathbb{E}_{X\sim P_X}[F(X)].$0, with $\mu \coloneqq \mathbb{E}_{X\sim P_X}[F(X)].$1 and $\mu \coloneqq \mathbb{E}_{X\sim P_X}[F(X)].$2. For each fold $\mu \coloneqq \mathbb{E}_{X\sim P_X}[F(X)].$3, a fold-specific surrogate is trained on the complement,

$\mu \coloneqq \mathbb{E}_{X\sim P_X}[F(X)].$4

where $\mu \coloneqq \mathbb{E}_{X\sim P_X}[F(X)].$5 excludes the labeled samples in fold $\mu \coloneqq \mathbb{E}_{X\sim P_X}[F(X)].$6 (Luo et al., 3 Jun 2026).

A fold-averaged surrogate is then defined on scenario-only data,

$\mu \coloneqq \mathbb{E}_{X\sim P_X}[F(X)].$7

and the cross-fitted estimator becomes

$\mu \coloneqq \mathbb{E}_{X\sim P_X}[F(X)].$8

Each residual is thus evaluated using a surrogate that did not train on that sample’s label, which preserves the unbiasedness argument (Luo et al., 3 Jun 2026).

For equal folds, the conditional variance decomposes as

$\mu \coloneqq \mathbb{E}_{X\sim P_X}[F(X)].$9

with

$X$0

A corresponding CLT-based confidence interval is

$X$1

where $X$2 is the empirical variance of cross-fitted surrogate predictions on scenario-only data and $X$3 is the empirical variance of the cross-fitted residuals on labeled data (Luo et al., 3 Jun 2026).

The assumptions underlying these guarantees are explicit. Target-domain labeled and scenario-only samples must be i.i.d. from the deployment distribution $X$4. For every residual term, the surrogate must be trained without the corresponding label. Approximate normality is assumed for the CLT-based confidence intervals. Notably, auxiliary data may be arbitrarily biased or shifted, because they influence the estimator only through the learned surrogate (Luo et al., 3 Jun 2026).

5. Empirical studies in autonomous driving and robot manipulation

The empirical evaluation covers autonomous driving and real-world manipulation, with metrics including mean final displacement error after 3 seconds, mean closed-loop distance to ground truth, and mean success rate. Across these studies, X4Val is compared with Monte Carlo, CPPI, CV, and CV_MCF baselines (Luo et al., 3 Jun 2026).

Setting	Data configuration	Reported outcome
Autonomous driving, geographic transfer	Germany target labels; 86,848 Germany scenario-only feature vectors; US auxiliary evaluation data	X4Val gives 15–20% variance reduction vs MC and outperforms CPPI and CV_MCF
Autonomous driving, iterative policy development	Five earlier policies with 1,000 paired samples each; current policy with 200 paired samples; 257 target scenario-only features	X4Val achieves 38.4% variance reduction vs MC
Real-world block stacking, cross-platform	5,000 ManiSkill rollouts; 100 real-world rollouts; 200 real-world scenario-only images	X4Val consistently reduces variance for all tested $X$5

In the geographic-transfer study, the policy is trained in the US and evaluated in Germany, with 3-second final displacement error in Germany as the target metric. X4Val uses the same auxiliary and target data as CV_MCF, but employs cross-fitting so that all target labeled data can contribute to both learning and estimation. The paper reports consistent 15–20% variance reduction relative to Monte Carlo across different sizes of Germany evaluation data, and notes that X4Val with a standard MSE objective performs somewhat worse than the variance-aligned version (Luo et al., 3 Jun 2026).

In the iterative-policy study, the target metric is closed-loop distance to ground truth at 3 seconds after engagement, an expensive closed-loop-simulation quantity. Five earlier policies provide 1,000 paired samples each, while the current policy provides only 200 paired samples. Open-loop final displacement error is available as a cheap control variate with correlation $X$6, making the classical CV baseline already strong. X4Val nevertheless delivers additional gains by amortizing information across earlier policies through the DeepSets-based meta-learning surrogate, and the reported variance reduction reaches 38.4% relative to Monte Carlo (Luo et al., 3 Jun 2026).

In the manipulation study, the task is block stacking with a visuomotor diffusion policy trained from 500 demonstrations in simulation and 100 demonstrations in the real world. Evaluation uses 5,000 ManiSkill rollouts, 100 real-world rollouts, and 200 real-world images of initial configurations without executing the policy. Because exact sim-real pairing of block poses is unavailable, this setting is structurally non-paired. X4Val pretrains on ManiSkill, finetunes on real-world labeled rollouts, and applies CPPI-style cross-fitting with $X$7. The reported result is consistent variance reduction for all $X$8, with improvement as $X$9 increases (Luo et al., 3 Jun 2026).

6. Interpretation, misconceptions, limitations, and extensions

X4Val is best understood as a prediction-powered control-variates framework specialized to non-paired, multi-domain evaluation. A common misconception is that auxiliary data must be distributionally matched to the deployment domain to be useful. The framework rejects that premise: auxiliary data may be substantially off-domain, because they are never inserted directly into the estimator of $F(X)$0. Their role is limited to improving the surrogate, while any residual surrogate bias is corrected on deployment-domain labels (Luo et al., 3 Jun 2026).

A second misconception is that variance reduction of this kind requires paired real and simulated outcomes. X4Val is explicitly constructed for the case in which such pairing is unavailable. This is the sense in which it generalizes classical control-variates methods: it retains the control-variate structure, but replaces a known or directly measurable auxiliary quantity with a learned neural surrogate whose expectation on the deployment distribution is estimated using cheap target-domain scenario-only samples (Luo et al., 3 Jun 2026).

The method has clear constraints. Target-domain labeled samples and target-domain scenario-only samples must both be i.i.d. from $F(X)$1; auxiliary data cannot replace target data in the estimator itself. Performance depends on the quality of the shared representation $F(X)$2, so poor embeddings or badly misaligned features can limit the correlation between $F(X)$3 and $F(X)$4. Cross-fitting entails multiple surrogate trainings and may be computationally expensive, especially in meta-learning regimes. The confidence intervals are CLT-based and can therefore be approximate when sample sizes are very small (Luo et al., 3 Jun 2026).

The framework also suggests several extensions. The paper identifies quantiles, tail-risk metrics, and conditional performance as natural targets beyond mean estimation. It also suggests using the surrogate for active validation, where scenarios are selected for labeling because they are expected to be especially informative. A plausible implication is that X4Val’s estimator construction is not confined to robotics: the same structure could be transplanted to other domains that have scarce target labels, abundant shifted auxiliary data, and cheap target-domain covariates, including the healthcare and recommender-system settings named in the paper (Luo et al., 3 Jun 2026).