CDFlow: Multi-Domain Methods and Applications

• CDFlow is a term covering distinct methods; in robotics, it redefines collision avoidance by modeling a distribution over minimal-distance configurations using Neural ODEs.
• In normalizing flows, CDFlow employs an architecture based on circulant and diagonal matrices to achieve efficient likelihood estimation and inversion.
• In power systems, CDFlow refers to a variant of Convex DistFlow that improves branch-flow relaxations for more accurate and efficient AC power-flow analysis.

CDFlow is a non-unique research name used for distinct methods in contemporary technical literature. In 2025, it denotes at least two unrelated contributions: a robot-motion-planning framework titled “CDFlow: Generative Gradient Flows for Configuration Space Distance Fields via Neural ODEs,” which learns a distribution over minimal-distance collision configurations in configuration space, and “CDFlow: Building Invertible Layers with Circulant and Diagonal Matrices,” a normalizing-flow architecture built from structured invertible linear layers. In power-systems usage, “CDFlow” can also appear as a name variant for Convex DistFlow (CDF) relaxations (Li et al., 17 Sep 2025, Feng et al., 29 Oct 2025, Coffrin et al., 2015).

1. Nomenclature and scope

Because the name is overloaded, the intended meaning of “CDFlow” depends on domain context. In robotics, it refers to a learned configuration-space distance-field framework. In generative modeling, it refers to Circulant-Diagonal Flow. In power systems, it may refer to Convex DistFlow usage around branch-flow relaxations.

Usage of “CDFlow”	Domain	Representative source
Configuration-space distance-field model via Neural ODEs	Robot motion planning	(Li et al., 17 Sep 2025)
Circulant-Diagonal Flow	Normalizing flows and density estimation	(Feng et al., 29 Oct 2025)
Convex DistFlow / CDF in some usage	AC power-flow relaxations	(Coffrin et al., 2015)

This naming overlap is a recurrent source of confusion. It is separate from similarly named methods such as DCFlow for unsupervised cross-modal optical flow, CrowdFlow for crowd optical-flow benchmarking, and ChangeFlow for remote-sensing change detection (Zhang et al., 29 Sep 2025, Schröder et al., 2018, Rolih et al., 14 May 2026).

2. Configuration-space CDFlow: problem reformulation

In robot motion planning, CDFlow begins from the distinction between task-space Signed Distance Fields (SDFs) and their configuration-space counterpart, the Configuration Space Distance Field (CDF). An SDF is defined as a continuous function $f_s:\mathbb{R}^3\to\mathbb{R}$ by

$$f_s(\mathbf{x})=\text{sign}(\mathbf{x})\min_{\mathbf{x}'\in\partial\Omega}\|\mathbf{x}-\mathbf{x}'\|_2,$$

with positive sign outside, negative inside, and zero on the boundary. The paper emphasizes that SDFs are differentiable almost everywhere and satisfy $\|\nabla f_s(\mathbf{x})\|_2=1$. A CDF transfers this construction to joint space: $$f_c(\mathbf{q})=\min_{\mathbf{q}'\in\mathcal Q_c}\|\mathbf{q}-\mathbf{q}'\|_2,$$ where $\mathbf{q}\in\mathcal Q$ and $\mathcal Q_c$ is the set of colliding configurations (Li et al., 17 Sep 2025).

The central critique is that classical CDFs conceal a uni-modal assumption. The $\min$ operator selects one nearest colliding configuration even when many minimal-distance collision configurations exist. In high-DoF robots, the set of minimal-distance collisions can be a manifold rather than a single point. The paper names the resulting failure mode gradient ambiguity: if several collision configurations are equally valid nearest solutions, the gradient obtained from any one of them is arbitrary. It also identifies geometric distortion, caused by sparse sampling of the collision boundary and optimization over that sparse set, which can oversmooth geometry and distort sharp high-dimensional boundary features (Li et al., 17 Sep 2025).

CDFlow replaces the pointwise nearest-collision view with a conditional distribution over all minimal-distance collision configurations,

$P(\mathbf{q}'\mid \mathbf{q}, \text{scene}),$

supported on $\mathcal Q^*(\mathbf q,\text{scene})$. Distance and gradient are then defined as expectations: $$f_{\text{flow}}(\mathbf{q})= \mathbb{E}_{\mathbf{q}'\sim P(\mathbf{q}'\mid \mathbf{q},\text{scene})} \big[\|\mathbf{q}-\mathbf{q}'\|_2\big],$$

$$f_s(\mathbf{x})=\text{sign}(\mathbf{x})\min_{\mathbf{x}'\in\partial\Omega}\|\mathbf{x}-\mathbf{x}'\|_2,$$0

This reformulation is the decisive conceptual move in the robotics meaning of CDFlow: the gradient is not the direction to one chosen nearest collision configuration, but the expected direction toward the entire minimal-distance collision distribution (Li et al., 17 Sep 2025).

3. Configuration-space CDFlow: generative model, data construction, and objective

To learn $$f_s(\mathbf{x})=\text{sign}(\mathbf{x})\min_{\mathbf{x}'\in\partial\Omega}\|\mathbf{x}-\mathbf{x}'\|_2,$$1, CDFlow uses a Neural ODE, specifically a conditional Continuous Normalizing Flow (CNF). The general dynamics are written as

$$f_s(\mathbf{x})=\text{sign}(\mathbf{x})\min_{\mathbf{x}'\in\partial\Omega}\|\mathbf{x}-\mathbf{x}'\|_2,$$2

with log-density evolution

$$f_s(\mathbf{x})=\text{sign}(\mathbf{x})\min_{\mathbf{x}'\in\partial\Omega}\|\mathbf{x}-\mathbf{x}'\|_2,$$3

The ODE is conditioned on query input $$f_s(\mathbf{x})=\text{sign}(\mathbf{x})\min_{\mathbf{x}'\in\partial\Omega}\|\mathbf{x}-\mathbf{x}'\|_2,$$4, containing the query configuration $$f_s(\mathbf{x})=\text{sign}(\mathbf{x})\min_{\mathbf{x}'\in\partial\Omega}\|\mathbf{x}-\mathbf{x}'\|_2,$$5 and scene information. A latent sample $$f_s(\mathbf{x})=\text{sign}(\mathbf{x})\min_{\mathbf{x}'\in\partial\Omega}\|\mathbf{x}-\mathbf{x}'\|_2,$$6 is integrated to $$f_s(\mathbf{x})=\text{sign}(\mathbf{x})\min_{\mathbf{x}'\in\partial\Omega}\|\mathbf{x}-\mathbf{x}'\|_2,$$7, producing a sample $$f_s(\mathbf{x})=\text{sign}(\mathbf{x})\min_{\mathbf{x}'\in\partial\Omega}\|\mathbf{x}-\mathbf{x}'\|_2,$$8 from the collision distribution. The unified network $$f_s(\mathbf{x})=\text{sign}(\mathbf{x})\min_{\mathbf{x}'\in\partial\Omega}\|\mathbf{x}-\mathbf{x}'\|_2,$$9 has two outputs for a given $\|\nabla f_s(\mathbf{x})\|_2=1$0: a generative mapping from latent noise $\|\nabla f_s(\mathbf{x})\|_2=1$1 to a collision sample $\|\nabla f_s(\mathbf{x})\|_2=1$2 via the CNF, and a direct regression of the scalar distance $\|\nabla f_s(\mathbf{x})\|_2=1$3. The gradient $\|\nabla f_s(\mathbf{x})\|_2=1$4 is obtained by automatic differentiation (Li et al., 17 Sep 2025).

Because the conditional collision distribution is not analytically available, the training set is built by adaptive refinement sampling. The first stage is global exploration: from random seeds in $\|\nabla f_s(\mathbf{x})\|_2=1$5, an optimizer such as L-BFGS solves a minimal-distance-to-boundary problem to discover diverse locally optimal collision configurations. The second stage is local refinement: around each high-quality solution $\|\nabla f_s(\mathbf{x})\|_2=1$6, the method samples initial points in a small ball $\|\nabla f_s(\mathbf{x})\|_2=1$7 and reruns local optimization, thereby densifying the sampled minimal-distance manifold. The resulting set $\|\nabla f_s(\mathbf{x})\|_2=1$8 is the empirical approximation of the support of $\|\nabla f_s(\mathbf{x})\|_2=1$9 (Li et al., 17 Sep 2025).

Training combines distribution learning and geometric supervision. The CNF is optimized with negative log-likelihood,

$$f_c(\mathbf{q})=\min_{\mathbf{q}'\in\mathcal Q_c}\|\mathbf{q}-\mathbf{q}'\|_2,$$0

The paper then defines purified target distance and gradient: $$f_c(\mathbf{q})=\min_{\mathbf{q}'\in\mathcal Q_c}\|\mathbf{q}-\mathbf{q}'\|_2,$$1

$$f_c(\mathbf{q})=\min_{\mathbf{q}'\in\mathcal Q_c}\|\mathbf{q}-\mathbf{q}'\|_2,$$2

The direct regressor is supervised by

$$f_c(\mathbf{q})=\min_{\mathbf{q}'\in\mathcal Q_c}\|\mathbf{q}-\mathbf{q}'\|_2,$$3

and the gradient field by the cosine-direction loss

$$f_c(\mathbf{q})=\min_{\mathbf{q}'\in\mathcal Q_c}\|\mathbf{q}-\mathbf{q}'\|_2,$$4

Two additional regularizers impose field structure: $$f_c(\mathbf{q})=\min_{\mathbf{q}'\in\mathcal Q_c}\|\mathbf{q}-\mathbf{q}'\|_2,$$5 The full objective is

$$f_c(\mathbf{q})=\min_{\mathbf{q}'\in\mathcal Q_c}\|\mathbf{q}-\mathbf{q}'\|_2,$$6

Architecturally, the CNF hidden dynamics network has four fully connected layers of width 1024 with ReLU activations, the regressive branch is a 5-layer MLP, and positional encoding is applied to the inputs. Training uses Adam with initial learning rate $$f_c(\mathbf{q})=\min_{\mathbf{q}'\in\mathcal Q_c}\|\mathbf{q}-\mathbf{q}'\|_2,$$7, decayed by a factor of $$f_c(\mathbf{q})=\min_{\mathbf{q}'\in\mathcal Q_c}\|\mathbf{q}-\mathbf{q}'\|_2,$$8, for $$f_c(\mathbf{q})=\min_{\mathbf{q}'\in\mathcal Q_c}\|\mathbf{q}-\mathbf{q}'\|_2,$$9 epochs; for each training sample, the implementation randomly samples $\mathbf{q}\in\mathcal Q$0 query points and $\mathbf{q}\in\mathcal Q$1 configurations. Total training time is reported as approximately 2 hours on a single NVIDIA RTX 4090 GPU (Li et al., 17 Sep 2025).

4. Configuration-space CDFlow: planning use, reported performance, and limitations

CDFlow is used as the geometric engine in both projection-based planning and optimization-based motion planning. In projection-style use, the planner repeatedly queries the current configuration $\mathbf{q}\in\mathcal Q$2 and scene, obtains a collision-related gradient from CDFlow, and updates the configuration along that direction until contact is established or a stopping criterion is reached. In optimization-based planning, gradients of SDF, CDF, or CDFlow are integrated into solvers such as IPOPT, QRQP, OSQP, and qpOASES (Li et al., 17 Sep 2025).

The reported gains are concentrated on high-DoF manipulation. In a PyBullet whole-body button-pressing task with a 7-DoF Franka Panda, a small sphere of radius $\mathbf{q}\in\mathcal Q$3 m is placed at 1000 random 3D positions. The SDF-based method achieves success rate $\mathbf{q}\in\mathcal Q$4, average steps $\mathbf{q}\in\mathcal Q$5, and query time $\mathbf{q}\in\mathcal Q$6 ms; the prior CDF method achieves success rate $\mathbf{q}\in\mathcal Q$7, average steps $\mathbf{q}\in\mathcal Q$8, and query time $\mathbf{q}\in\mathcal Q$9 ms; CDFlow achieves success rate $\mathcal Q_c$0, average steps $\mathcal Q_c$1, and query time $\mathcal Q_c$2 ms. In MotionBenchMaker with a 7-DoF Franka over 10 tasks with 100 problems each, the paper reports that with IPOPT, success increases from $\mathcal Q_c$3 with SDF and $\mathcal Q_c$4 with CDF to $\mathcal Q_c$5 with CDFlow, while average joint-space tracking error drops from $\mathcal Q_c$6 and $\mathcal Q_c$7 rad to $\mathcal Q_c$8 rad, and optimization steps drop from 237 and 134 to 89. The paper summarizes an average $\mathcal Q_c$9 success rate across the benchmark (Li et al., 17 Sep 2025).

Setting	Baselines	Reported CDFlow result
PyBullet button pressing	SDF: $\min$0, 4.1, 2.13 ms; CDF: $\min$1, 2.3, 1.57 ms	$\min$2, 1.2, 0.81 ms
MotionBenchMaker with IPOPT	SDF: $\min$3, $\min$4 rad, 237; CDF: $\min$5, $\min$6 rad, 134	$\min$7, $\min$8 rad, 89
MotionBenchMaker average	Multiple solvers	$\min$9 success

The paper also emphasizes limitations. The method requires an offline data-generation phase for each new scene, which the authors note can be time-consuming. The most accurate gradient relies on Monte Carlo sampling from the learned distribution, and accuracy depends on the number of samples. Scene conditioning and scene-specific offline datasets imply environment dependence. The method is not evaluated for online adaptation to truly dynamic environments; this is presented as future work. Detailed ablation tables isolating multimodal modeling, Neural ODE design, and adaptive refinement sampling are not provided (Li et al., 17 Sep 2025).

5. Circulant-Diagonal Flow: CDFlow in normalizing flows

A second, unrelated use of the name denotes Circulant-Diagonal Flow, a normalizing-flow model built around an invertible linear layer represented as an alternating product of diagonal and circulant matrices. The weight matrix is written as

$P(\mathbf{q}'\mid \mathbf{q}, \text{scene}),$0

equivalently

$P(\mathbf{q}'\mid \mathbf{q}, \text{scene}),$1

The motivation is classical for flow models: efficient likelihood estimation and sampling require linear layers that are expressive but also permit cheap Jacobian log-determinants and inverses. The paper states that this decomposition reduces parameter complexity from $P(\mathbf{q}'\mid \mathbf{q}, \text{scene}),$2 to $P(\mathbf{q}'\mid \mathbf{q}, \text{scene}),$3 using $P(\mathbf{q}'\mid \mathbf{q}, \text{scene}),$4 diagonal matrices and $P(\mathbf{q}'\mid \mathbf{q}, \text{scene}),$5 circulant matrices, while reducing inversion and log-determinant costs to $P(\mathbf{q}'\mid \mathbf{q}, \text{scene}),$6 and $P(\mathbf{q}'\mid \mathbf{q}, \text{scene}),$7, respectively (Feng et al., 29 Oct 2025).

The key algebraic fact is Fourier diagonalization of a circulant matrix: $P(\mathbf{q}'\mid \mathbf{q}, \text{scene}),$8 This yields simple determinant and inverse formulas. For the full layer,

$P(\mathbf{q}'\mid \mathbf{q}, \text{scene}),$9

so the log-determinant becomes a sum over diagonal entries and Fourier-domain eigenvalues. The inverse has reversed factor order,

$\mathcal Q^*(\mathbf q,\text{scene})$0

The paper integrates this layer into a Glow-style architecture with ActNorm, CD-Convolution, and coupling layers inside a multi-scale design. In the reported configuration, increasing the number of factors $\mathcal Q^*(\mathbf q,\text{scene})$1 slightly improves expressiveness but increases parameter count and latency, so the main experiments use $\mathcal Q^*(\mathbf q,\text{scene})$2 (Feng et al., 29 Oct 2025).

Empirically, the model is evaluated on CIFAR-10, ImageNet $\mathcal Q^*(\mathbf q,\text{scene})$3, CIFAR-100, and Galaxy images. The reported bits-per-dimension values are $\mathcal Q^*(\mathbf q,\text{scene})$4 on CIFAR-10, $\mathcal Q^*(\mathbf q,\text{scene})$5 on ImageNet $\mathcal Q^*(\mathbf q,\text{scene})$6, and $\mathcal Q^*(\mathbf q,\text{scene})$7 on CIFAR-100, improving over Glow, Emerging, Woodbury, and ButterflyFlow in the corresponding tables. On the Galaxy dataset, the paper states that CDFlow outperforms all other models and achieves a $\mathcal Q^*(\mathbf q,\text{scene})$8 improvement in BPD compared to Glow. The abstract additionally reports that when the number of channels is 96, the method speeds up matrix inverse by $\mathcal Q^*(\mathbf q,\text{scene})$9 and matrix log-determinant by $$f_{\text{flow}}(\mathbf{q})= \mathbb{E}_{\mathbf{q}'\sim P(\mathbf{q}'\mid \mathbf{q},\text{scene})} \big[\|\mathbf{q}-\mathbf{q}'\|_2\big],$$0 relative to a general dense matrix (Feng et al., 29 Oct 2025).

This usage of CDFlow therefore belongs to structured invertible linear algebra rather than robotics. Its “flow” is a normalizing flow over data distributions, not a geometric gradient field in robot configuration space.

6. Related names, misconceptions, and citation practice

A common misconception is that “CDFlow” identifies a single established architecture. The documented literature instead supports at least two unrelated 2025 meanings, plus an older power-systems usage in which Convex DistFlow or CDF can be called CDFlow in some usage (Coffrin et al., 2015). A second misconception is that similarly named methods are variants of the same framework. The available sources do not support that reading: DCFlow is an unsupervised cross-modal flow-estimation framework, CrowdFlow is a crowd optical-flow benchmark, and ChangeFlow is a latent rectified-flow model for remote-sensing change detection (Zhang et al., 29 Sep 2025, Schröder et al., 2018, Rolih et al., 14 May 2026).

Within robotics, another misconception is that CDFlow merely predicts a better nearest collision point. The paper’s stated contribution is stronger and more specific: it redefines the problem from finding a single nearest point to modeling the distribution of minimal-distance collision configurations, and then derives distance and gradient as expectations over that distribution (Li et al., 17 Sep 2025). Within generative modeling, Circulant-Diagonal Flow is not a generic dense invertible layer replacement; it is a specific alternating product of diagonal and circulant matrices with FFT-enabled inverse and determinant computations (Feng et al., 29 Oct 2025).

This suggests that scholarly references to “CDFlow” benefit from explicit disambiguation by title, domain, or arXiv identifier. In current usage, the term can denote a configuration-space geometric representation for collision-aware planning, a structured normalizing-flow architecture, or, in some power-systems contexts, a DistFlow-related relaxation.