KDPE: Robotics and Semiparametric Estimation

• KDPE is a polysemous term that includes an inference-time KDE wrapper in robotics and an RKHS-based debiasing framework in semiparametric statistics.
• In robotics, KDPE applies manifold-aware kernel density estimation on candidate trajectories generated by diffusion policies to filter out unreliable actions and improve task robustness.
• In semiparametric estimation, KDPE employs recursive RKHS-based likelihood updates to debias plug-in estimators, achieving regular, asymptotically efficient inference without explicit influence functions.

KDPE is an acronym used for distinct research constructs in contemporary arXiv literature. In robotics, it denotes “KDPE: A Kernel Density Estimation Strategy for Diffusion Policy Trajectory Selection”, an inference-time method that filters candidate trajectories generated by Diffusion Policy using a manifold-aware kernel density estimator over robot actions (Rosasco et al., 14 Aug 2025). In semiparametric statistics, it denotes Kernel Debiased Plug-in Estimation, an RKHS-based debiasing framework for plug-in estimators that aims at regular, asymptotically linear, semiparametrically efficient estimation without explicit influence-function derivation (Cho et al., 2023, Chen et al., 9 Mar 2026). A plausible implication is that “KDPE” should be treated as a domain-dependent term rather than a single established concept.

1. KDPE in robot policy learning

In robot imitation learning, KDPE is introduced in the standard visuomotor behavior cloning setting, where a policy maps the current observation to an action trajectory. The method is designed as an inference-time enhancement for Diffusion Policy (DP), which models the policy as a denoising diffusion probabilistic model and generates an action trajectory $\mathbf{A} \in \mathbb{R}^{T \times D}$ by iterative denoising from $\mathbf{A}^K \sim \mathcal{N}(0,I)$ (Rosasco et al., 14 Aug 2025). The diffusion update is written as

$$\mathbf{A}^{k-1} = \alpha\big(\mathbf{A}^k - \gamma \epsilon_{\theta}(\mathbf{A}^k, k) + \mathcal{N}(0, \sigma^2 I)\big).$$

The motivation is explicitly tied to two drawbacks of Diffusion Policy. First, the stochasticity of the denoising process can highly impact the quality of the generated action trajectory. Second, because DP is a supervised learning approach, it can learn data outliers from the training dataset and reproduce them at test time (Rosasco et al., 14 Aug 2025). The reported downstream issues include out-of-distribution trajectories, unreliable behavior across resets, and sensitivity to mild domain shift.

KDPE addresses these issues by wrapping a trained Diffusion Policy and statistically filtering the trajectories it generates. At each decision step, DP samples multiple candidate trajectories in parallel, KDE is computed over the final action of each candidate, and only the trajectory whose last action is most in-distribution relative to the demonstrations is executed (Rosasco et al., 14 Aug 2025). No changes to DP’s architecture or training procedure are required; the method is purely test-time.

2. Kernel density formulation for trajectory selection

KDPE in robotics samples $N$ candidate trajectories

$$\{\mathbf{A}_i\}_{i=1}^N,\quad \mathbf{A}_i \in \mathbb{R}^{T \times D},$$

extracts the last action from each trajectory,

$\mathbf{a}_i = \mathbf{A}_i[T],$

and estimates a density over these actions by kernel density estimation (Rosasco et al., 14 Aug 2025). The estimated density at an arbitrary action $\tilde{\mathbf{a}}$ is

$$\rho(\tilde{\mathbf{a}}) = \frac{1}{N}\sum_{i=1}^{N} k(\tilde{\mathbf{a}}, \mathbf{a}_i).$$

The selected trajectory is

$i^\star = \arg\max_i \rho(\mathbf{a}_i).$

The action representation is mixed-geometry: end-effector Cartesian position $\mathbf{t}\in\mathbb{R}^3$, orientation $\mathbf{A}^K \sim \mathcal{N}(0,I)$0, and gripper aperture $\mathbf{A}^K \sim \mathcal{N}(0,I)$1, so that

$\mathbf{A}^K \sim \mathcal{N}(0,I)$2

The experiments use action dimension $\mathbf{A}^K \sim \mathcal{N}(0,I)$3, corresponding to $\mathbf{A}^K \sim \mathcal{N}(0,I)$4 position coordinates, a $\mathbf{A}^K \sim \mathcal{N}(0,I)$5D orientation representation, and one gripper scalar, while the kernel itself is formulated in terms of $\mathbf{A}^K \sim \mathcal{N}(0,I)$6 matrices for clarity (Rosasco et al., 14 Aug 2025).

The kernel is described as a manifold-aware kernel adapted from Barfoot and Furgale. Its difference vector is

$\mathbf{A}^K \sim \mathcal{N}(0,I)$7

with diagonal bandwidth matrix

$\mathbf{A}^K \sim \mathcal{N}(0,I)$8

The rotational term uses the Lie-group logarithm and vee operator so that distances reflect geodesic angles on $\mathbf{A}^K \sim \mathcal{N}(0,I)$9 rather than naive component-wise differences (Rosasco et al., 14 Aug 2025). This is the sense in which the kernel is “manifold-aware”: Euclidean coordinates are handled in Euclidean form, whereas orientations are compared via tangent-space coordinates induced by the log map.

The density score of a sampled action is

$$\mathbf{A}^{k-1} = \alpha\big(\mathbf{A}^k - \gamma \epsilon_{\theta}(\mathbf{A}^k, k) + \mathcal{N}(0, \sigma^2 I)\big).$$0

so high-density actions are those lying in well-populated local clusters, while isolated samples receive low density. The paper also evaluates a stress-test variant, KDPE-OOD, which deliberately selects the lowest-density action; it performs substantially worse, supporting the interpretation of low-density trajectories as harmful outliers (Rosasco et al., 14 Aug 2025).

3. Integration with Diffusion Policy and empirical findings

KDPE is implemented online rather than as a global offline density model. The KDE is fitted over the small sample of actions generated at each control step, so there is no separate density-model training phase (Rosasco et al., 14 Aug 2025). The reported bandwidths are

$$\mathbf{A}^{k-1} = \alpha\big(\mathbf{A}^k - \gamma \epsilon_{\theta}(\mathbf{A}^k, k) + \mathcal{N}(0, \sigma^2 I)\big).$$1

These were tuned using a trajectory visualizer so that multiple clusters appear where DP outputs distinct modes.

The experiments use $$\mathbf{A}^{k-1} = \alpha\big(\mathbf{A}^k - \gamma \epsilon_{\theta}(\mathbf{A}^k, k) + \mathcal{N}(0, \sigma^2 I)\big).$$2 sampled trajectories per decision step, execution horizon $$\mathbf{A}^{k-1} = \alpha\big(\mathbf{A}^k - \gamma \epsilon_{\theta}(\mathbf{A}^k, k) + \mathcal{N}(0, \sigma^2 I)\big).$$3 actions, and $$\mathbf{A}^{k-1} = \alpha\big(\mathbf{A}^k - \gamma \epsilon_{\theta}(\mathbf{A}^k, k) + \mathcal{N}(0, \sigma^2 I)\big).$$4k training steps for the underlying DP models (Rosasco et al., 14 Aug 2025). Both CNN-based and Transformer-based Diffusion Policy backbones are used, denoted DP-C and DP-T. KDPE-C and KDPE-T are the corresponding inference-time selection wrappers.

The simulated evaluation covers seven tasks drawn from RoboMimic and MimicGen: Lift, Can, Square, ToolHang, Coffee, Stack Three, and Three Piece Assembly. The average success rates reported over all seven tasks are $$\mathbf{A}^{k-1} = \alpha\big(\mathbf{A}^k - \gamma \epsilon_{\theta}(\mathbf{A}^k, k) + \mathcal{N}(0, \sigma^2 I)\big).$$5 for DP-C versus $$\mathbf{A}^{k-1} = \alpha\big(\mathbf{A}^k - \gamma \epsilon_{\theta}(\mathbf{A}^k, k) + \mathcal{N}(0, \sigma^2 I)\big).$$6 for KDPE-C, and $$\mathbf{A}^{k-1} = \alpha\big(\mathbf{A}^k - \gamma \epsilon_{\theta}(\mathbf{A}^k, k) + \mathcal{N}(0, \sigma^2 I)\big).$$7 for DP-T versus $$\mathbf{A}^{k-1} = \alpha\big(\mathbf{A}^k - \gamma \epsilon_{\theta}(\mathbf{A}^k, k) + \mathcal{N}(0, \sigma^2 I)\big).$$8 for KDPE-T (Rosasco et al., 14 Aug 2025). On perturbed environments produced by a mild color shift, the average success rates are $$\mathbf{A}^{k-1} = \alpha\big(\mathbf{A}^k - \gamma \epsilon_{\theta}(\mathbf{A}^k, k) + \mathcal{N}(0, \sigma^2 I)\big).$$9 for DP-C versus $N$0 for KDPE-C, and $N$1 for DP-T versus $N$2 for KDPE-T. On the precision-critical ToolHang task, the gains are larger: DP-C $N$3 versus KDPE-C $N$4, and DP-T $N$5 versus KDPE-T $N$6 (Rosasco et al., 14 Aug 2025).

The real-robot evaluation uses a Franka Emika Panda with a Robotiq 2f-85 gripper, wrist and external RGB cameras, and teleoperated data collection with Meta Quest 3. Reported success rates are $N$7 for DP-C versus $N$8 for KDPE-C on PickPlush, $N$9 versus $$\{\mathbf{A}_i\}_{i=1}^N,\quad \mathbf{A}_i \in \mathbb{R}^{T \times D},$$0 on PickSponge, $$\{\mathbf{A}_i\}_{i=1}^N,\quad \mathbf{A}_i \in \mathbb{R}^{T \times D},$$1 versus $$\{\mathbf{A}_i\}_{i=1}^N,\quad \mathbf{A}_i \in \mathbb{R}^{T \times D},$$2 on CubeSort, and $$\{\mathbf{A}_i\}_{i=1}^N,\quad \mathbf{A}_i \in \mathbb{R}^{T \times D},$$3 versus $$\{\mathbf{A}_i\}_{i=1}^N,\quad \mathbf{A}_i \in \mathbb{R}^{T \times D},$$4 on CoffeeMaking (Rosasco et al., 14 Aug 2025). The paper additionally states that KDPE yields smoother and more consistent behavior.

Test-time overhead is measured on an RTX 3080. Baseline DP-C with single-trajectory DDIM sampling takes about $$\{\mathbf{A}_i\}_{i=1}^N,\quad \mathbf{A}_i \in \mathbb{R}^{T \times D},$$5 ms per inference step; generating $$\{\mathbf{A}_i\}_{i=1}^N,\quad \mathbf{A}_i \in \mathbb{R}^{T \times D},$$6 trajectories instead of $$\{\mathbf{A}_i\}_{i=1}^N,\quad \mathbf{A}_i \in \mathbb{R}^{T \times D},$$7 adds about $$\{\mathbf{A}_i\}_{i=1}^N,\quad \mathbf{A}_i \in \mathbb{R}^{T \times D},$$8 ms; KDE over the $$\{\mathbf{A}_i\}_{i=1}^N,\quad \mathbf{A}_i \in \mathbb{R}^{T \times D},$$9 final actions adds about $\mathbf{a}_i = \mathbf{A}_i[T],$0 ms; total KDPE-C inference is about $\mathbf{a}_i = \mathbf{A}_i[T],$1 ms, corresponding to about $\mathbf{a}_i = \mathbf{A}_i[T],$2 Hz versus about $\mathbf{a}_i = \mathbf{A}_i[T],$3 Hz for original DP-C (Rosasco et al., 14 Aug 2025). A trajectory-level variant, Tr-KDPE, is more expensive and less effective on average, which the paper attributes to the curse of dimensionality in higher-dimensional KDE.

4. KDPE as Kernel Debiased Plug-in Estimation

In semiparametric and nonparametric statistics, KDPE stands for Kernel Debiased Plug-in Estimation. The setting is i.i.d. data $\mathbf{a}_i = \mathbf{A}_i[T],$4 in a nonparametric or semiparametric model, together with one or more pathwise differentiable target parameters $\mathbf{a}_i = \mathbf{A}_i[T],$5 or $\mathbf{a}_i = \mathbf{A}_i[T],$6 (Cho et al., 2023, Chen et al., 9 Mar 2026). The basic plug-in estimator has the form

$\mathbf{a}_i = \mathbf{A}_i[T],$7

where $\mathbf{a}_i = \mathbf{A}_i[T],$8 is an estimated distribution.

The central problem is plug-in bias. For pathwise differentiable targets, the canonical expansion includes a term of the form $\mathbf{a}_i = \mathbf{A}_i[T],$9, where $\tilde{\mathbf{a}}$0 is the canonical gradient or efficient influence function. Traditional efficient procedures such as one-step estimators, double machine learning, and TMLE remove this first-order bias by using explicit influence-function calculations (Cho et al., 2023). KDPE is proposed precisely to avoid explicit influence-function derivation while retaining semiparametric efficiency under regularity conditions.

The 2023 formulation places KDPE within the TMLE framework. Starting from a pre-estimate $\tilde{\mathbf{a}}$1, it performs recursive regularized likelihood-maximization updates over an RKHS-based fluctuation model

$\tilde{\mathbf{a}}$2

where $\tilde{\mathbf{a}}$3 is a mean-zero RKHS at the current distribution $\tilde{\mathbf{a}}$4 (Cho et al., 2023). The optimization step is

$\tilde{\mathbf{a}}$5

subject to membership in the fluctuation model.

The representer theorem reduces this infinite-dimensional problem to a finite-dimensional optimization over kernel sections evaluated at the observed sample. The resulting recursion produces a limit distribution $\tilde{\mathbf{a}}$6 such that

$\tilde{\mathbf{a}}$7

and the final KDPE estimator of any target parameter is

$\tilde{\mathbf{a}}$8

(Cho et al., 2023). Because the underlying RKHS is assumed universal, the mean-zero RKHS is dense in $\tilde{\mathbf{a}}$9, so efficient influence functions can be approximated without being explicitly written down. This is the basis for the claim that KDPE simultaneously debiases all pathwise differentiable target parameters satisfying the regularity conditions.

5. ULFS-KDPE and the universal least favorable submodel

A later development, ULFS-KDPE, reformulates KDPE around the universal least favorable submodel (ULFS) and gives a functional-analytic ODE treatment (Chen et al., 9 Mar 2026). The classical least favorable submodel is local: at a fixed $$\rho(\tilde{\mathbf{a}}) = \frac{1}{N}\sum_{i=1}^{N} k(\tilde{\mathbf{a}}, \mathbf{a}_i).$$0, its score at the starting point equals the canonical gradient. The ULFS generalizes this to a global path $$\rho(\tilde{\mathbf{a}}) = \frac{1}{N}\sum_{i=1}^{N} k(\tilde{\mathbf{a}}, \mathbf{a}_i).$$1 such that, at every point,

$$\rho(\tilde{\mathbf{a}}) = \frac{1}{N}\sum_{i=1}^{N} k(\tilde{\mathbf{a}}, \mathbf{a}_i).$$2

This is a nonlinear ODE on densities.

ULFS-KDPE does not use the efficient influence function analytically. Instead, it defines an RKHS-based direction $$\rho(\tilde{\mathbf{a}}) = \frac{1}{N}\sum_{i=1}^{N} k(\tilde{\mathbf{a}}, \mathbf{a}_i).$$3 and studies the ODE

$$\rho(\tilde{\mathbf{a}}) = \frac{1}{N}\sum_{i=1}^{N} k(\tilde{\mathbf{a}}, \mathbf{a}_i).$$4

The empirical quantity controlling the flow is

$$\rho(\tilde{\mathbf{a}}) = \frac{1}{N}\sum_{i=1}^{N} k(\tilde{\mathbf{a}}, \mathbf{a}_i).$$5

while the empirical log-likelihood is

$$\rho(\tilde{\mathbf{a}}) = \frac{1}{N}\sum_{i=1}^{N} k(\tilde{\mathbf{a}}, \mathbf{a}_i).$$6

The paper shows that

$$\rho(\tilde{\mathbf{a}}) = \frac{1}{N}\sum_{i=1}^{N} k(\tilde{\mathbf{a}}, \mathbf{a}_i).$$7

so the flow is nondecreasing in empirical log-likelihood, and stationarity occurs exactly when the empirical mean embedding vanishes (Chen et al., 9 Mar 2026).

The RKHS machinery is explicit. For kernel $$\rho(\tilde{\mathbf{a}}) = \frac{1}{N}\sum_{i=1}^{N} k(\tilde{\mathbf{a}}, \mathbf{a}_i).$$8, the mean-zero RKHS at $$\rho(\tilde{\mathbf{a}}) = \frac{1}{N}\sum_{i=1}^{N} k(\tilde{\mathbf{a}}, \mathbf{a}_i).$$9 is

$i^\star = \arg\max_i \rho(\mathbf{a}_i).$0

with centered kernel

$i^\star = \arg\max_i \rho(\mathbf{a}_i).$1

The empirical mean embedding is

$i^\star = \arg\max_i \rho(\mathbf{a}_i).$2

and the debiasing direction is

$i^\star = \arg\max_i \rho(\mathbf{a}_i).$3

This gives a finite-matrix implementation via the centered Gram matrix $i^\star = \arg\max_i \rho(\mathbf{a}_i).$4 and coefficient vector

$i^\star = \arg\max_i \rho(\mathbf{a}_i).$5

(Chen et al., 9 Mar 2026).

The paper proves existence, uniqueness, stability, and finite-time convergence of the empirical score along this flow under bounded smooth-kernel assumptions and Hölder regularity of the density class (Chen et al., 9 Mar 2026). It further states that, under standard regularity conditions, the resulting plug-in estimator is regular, asymptotically linear, and semiparametrically efficient for a broad class of pathwise differentiable parameters. Simulation studies with $i^\star = \arg\max_i \rho(\mathbf{a}_i).$6 Monte Carlo replications at $i^\star = \arg\max_i \rho(\mathbf{a}_i).$7 compare ULFS-KDPE, original KDPE, TMLE, and one-step TMLE on ATE, risk ratio, and odds ratio. The reported findings are that ULFS-KDPE generally has lower bias than original KDPE and TMLE, comparable or smaller variance, and favorable RMSE, especially in the positivity-challenged DGP2 setting (Chen et al., 9 Mar 2026).

6. Conceptual contrasts, related usage, and ambiguity

The two principal senses of KDPE differ in objective, mathematical substrate, and computational role.

Usage	Domain	Core mechanism
KDPE (Rosasco et al., 14 Aug 2025)	Robot imitation learning	Online KDE over sampled robot actions for trajectory selection
KDPE (Cho et al., 2023)	Semiparametric statistics	RKHS-based likelihood targeting to remove plug-in bias
ULFS-KDPE (Chen et al., 9 Mar 2026)	Semiparametric statistics	ODE-based RKHS debiasing flow along a universal least favorable submodel

In robotics, KDPE is an inference-time wrapper around a trained generative policy. It selects among sampled trajectories and leaves the underlying policy unchanged (Rosasco et al., 14 Aug 2025). In semiparametric statistics, KDPE is a distributional debiasing procedure that modifies an estimated data-generating distribution so that score equations are approximately solved for a rich function class (Cho et al., 2023). The commonality is only at a high level: both use kernels and both perform a selection or updating step based on a notion of density or score geometry. This suggests a lexical overlap rather than a shared methodology.

The data block also includes a PDE-discovery paper, “KO-PDE-IDENT” (Thanasutives et al., 26 May 2026). That paper explicitly states that the query “KDPE” is not an acronym used in the paper, though it plausibly refers to a broader knockoff-based PDE-discovery paradigm. The framework itself is named KO-PDE-IDENT and combines model-X knockoff filters, SHAP-based feature statistics, recursive feature elimination, and multi-criteria decision-making for discovering parsimonious governing PDEs with finite-sample FDR control (Thanasutives et al., 26 May 2026). Accordingly, it is related only through acronym ambiguity, not through a named KDPE method.

A common misconception would be to treat KDPE as a single canonical technique across fields. The literature summarized here does not support that interpretation. Instead, KDPE is a polysemous acronym whose meaning must be resolved from disciplinary context.

7. Significance and research directions

The robotics formulation positions KDPE as a lightweight mechanism for turning the stochasticity of Diffusion Policy into a multi-hypothesis controller with a statistical filter (Rosasco et al., 14 Aug 2025). The reported benefits are strongest on precision tasks, under lower-quality demonstrations, and under mild visual perturbations. The paper identifies several limitations: KDPE can only select among trajectories that DP already generates; it may reduce beneficial stochasticity when unusual recovery behavior would help; and higher-dimensional trajectory-level KDE suffers from the curse of dimensionality (Rosasco et al., 14 Aug 2025). Suggested directions include guided denoising, extension to higher-dimensional action spaces such as bimanual manipulation and dexterous hands, and application to other generative robot policies.

The statistical KDPE line positions RKHS-based debiasing as an alternative to influence-function-specific targeting. The 2023 paper emphasizes simultaneous debiasing of many pathwise differentiable parameters from a single debiased distribution (Cho et al., 2023), while the 2026 ULFS-KDPE paper adds an ODE-based foundation, Lyapunov structure, finite-time stopping guarantees, and improved numerical stability in simulation (Chen et al., 9 Mar 2026). The main limitations stated in these works are computational cost from kernel Gram matrices, sensitivity to kernel and bandwidth choices, smoothness assumptions on densities, and the need for tuned stopping criteria. Proposed extensions include random-feature or Nyström approximations, longitudinal and time-to-event causal settings, higher-order targeting, and adaptive stopping rules with explicit inference guarantees (Cho et al., 2023, Chen et al., 9 Mar 2026).

Taken together, the literature presents KDPE as an acronym attached to technically mature but unrelated research programs: one centered on kernel density estimation for diffusion-policy trajectory selection in robotics, the other on kernel debiased plug-in estimation in semiparametric inference. Any use of the term in isolation is therefore incomplete unless the surrounding field is specified.