ManiFlow: Manifold Flow & Optimal Transport

• ManiFlow is a class of techniques that leverage flow-based learning, manifold representations, and optimal transport to model high-dimensional dynamical phenomena with robust geometric fidelity.
• It applies to real-world tasks such as robotic manipulation, shape reconstruction, and trajectory inference by encoding lower-dimensional manifold constraints and enforcing physical consistency.
• Techniques like flow matching, consistency training, and neural ODEs enhance performance, scalability, and efficiency across simulation and real-world applications.

ManiFlow refers to a class of methods and frameworks that leverage flow-based learning, manifold representations, and optimal transport to model, infer, and control high-dimensional dynamical phenomena or actions, with prominent applications in robotics, shape modeling, trajectory generation, and pose estimation. Unlike conventional approaches that operate in unstructured ambient space, ManiFlow methods intrinsically encode lower-dimensional manifold constraints, exploit flow matching to connect data-driven distributions, and impose geometrically faithful or physically meaningful consistency during sample generation or policy inference. Research in this area includes meshfree simulation of PDEs on curved surfaces, generative models that respect manifold topology, flow-based trajectory inference, and policy learning that scales across sensor, language, and embodiment modalities.

1. Manifold Representation and Flow-Based Modeling

A central theme in ManiFlow techniques is the development of representations and generative models that explicitly or implicitly respect the manifold support of the underlying data. In the context of generative modeling (e.g., normalizing flows), this is achieved by parameterizing a diffeomorphic mapping $f: \mathbb{R}^d \to \mathcal{M}$ that transforms samples from a simple base distribution (typically Gaussian) into complex points lying on a target manifold $\mathcal{M}$ embedded in higher dimensions (Postels et al., 2022).

For fluid simulation on manifolds, a meshfree Lagrangian framework is introduced in which a cloud of points discretizes a curved surface $M\subset\mathbb{R}^3$, and manifold constraint is maintained through tangential particle advection and geometric projection without explicit surface parameterization (Suchde, 2020). For pose and trajectory inference, manifolds such as SO(3) (for rotations) or learned latent spaces for trajectories are used as domains for conditional normalizing flows or Neural ODEs (Sengupta et al., 2023, Huguet et al., 2022).

The table below summarizes ManiFlow representations across applications:

Context	Manifold	Representation
Shape modeling	Low-dimensional in $\mathbb{R}^D$	Normalizing flow, explicit density
Fluid simulation	Curved surface $M\subset\mathbb{R}^3$	Particle cloud, tangential operators
Human pose/rotation	SO(3) manifold	Normalizing flow on 𝖘𝖔(3) + exp map
Trajectory planning	Learned nonlinear latent manifold	Autoencoder + flow matching

Preserving the manifold structure is crucial for enabling smooth interpolation, accurate density estimation, and generation of physically or statistically valid samples.

2. Flow Matching, Consistency Training, and Optimal Transport

ManiFlow methods frequently employ flow matching—a conditional deep generative modeling approach that learns vector fields aligning simple and complex distributions across a controlled transformation (e.g., from noise to data or between data snapshots) (Lee et al., 29 Jul 2024, Yan et al., 1 Sep 2025). The key training objective is a flow matching loss of the form: $$\mathcal{L}_{\mathrm{FM}}(\theta) = \mathbb{E}_{x_0,x_1\sim \mathcal{D},\,t\sim\mathcal{U}[0,1]} \left\|v_\theta(x_t, t) - (x_1 - x_0)\right\|^2$$ where $x_t = (1-t)x_0 + t x_1$ and $v_\theta$ predicts the instantaneous velocity needed to connect $x_0$ and $x_1$ at fraction $t$.

Consistency flow training introduces an additional objective to align velocities along ODE-integration trajectories, ensuring that iteratively generated samples remain self-consistent and that fast (1–2 step) inference remains feasible.

In population and trajectory inference contexts (e.g., MIOFlow), the system is regularized to align with dynamic optimal transport, minimizing the integral energy of paths between marginal distributions under manifold-aware ground metrics (Huguet et al., 2022). The induced trajectory then corresponds to a minimal energy manifold geodesic under the Wasserstein-2 metric. These techniques are realized in Neural ODEs operating in learned latent spaces that match geodesic structure through auxiliary autoencoder losses.

3. Architecture and Multimodal Conditioning

Contemporary ManiFlow policies and generative models employ advanced neural architectures to couple multimodal observations—images, language, proprioception—with flow-based action or trajectory generation. The DiT-X architecture (Yan et al., 1 Sep 2025) exemplifies this direction, combining the following key features:

• Adaptive Cross-Attention: Action tokens selectively attend to processed embeddings from visual, language, and proprioceptive sources, enhancing fine-grained fusion and retrieval of task-relevant features.
• AdaLN-Zero Conditioning: Timesteps and low-dimensional controls are injected via adaptive layer normalization, enabling the network to modulate semantic and spatial features based on task phase and embodiment state.
• Transformer Dynamics: For sequence modeling, transformers with causal masks model temporal dependencies, while learnable query tokens fetch information relevant for downstream modules such as action generation or future image synthesis (He et al., 14 Feb 2025).

This flexible multimodal integration is fundamental to the ability of ManiFlow systems to generalize and scale across tasks and input scenarios, including real-world robot manipulation under natural language commands, 3D vision, and dexterous multi-arm contexts.

4. Applications: Manipulation, Shape Reconstruction, Pose and Trajectory Inference

ManiFlow frameworks have demonstrated significant empirical advances in several domains:

• Robotic Manipulation: ManiFlow policies generate dexterous multi-joint actions from images, language, and proprioception, showing strong real-world performance on single-arm, bimanual, and humanoid robots. Policies trained with flow matching and consistency objectives nearly double success rates on challenging tasks and maintain high robustness to object and background changes (Yan et al., 1 Sep 2025).
• Shape Modeling and Reconstruction: For point cloud modeling, ManiFlow employs normalizing flows to generate manifold-constrained samples and guide Poisson surface reconstruction using explicit likelihoods and normal consistency (Postels et al., 2022).
• Human Pose Estimation: HuManiFlow leverages SO(3)-respecting flows with autoregressive factorization along kinematic trees to generate accurate, diverse 3D poses consistent with 2D observations. This eliminates distribution collapse and improves inference under uncertainty (Sengupta et al., 2023).
• Trajectory and Population Inference: MIOFlow and MMFP use (Neural) ODEs and flow matching in latent manifolds for interpolating biological or robot trajectories, with constraints from optimal transport and geodesic consistency. These approaches outperform normalizing flows and Schrödinger bridge-based methods in capturing nonlinear, manifold-constrained dynamics and population shifts (Huguet et al., 2022, Lee et al., 29 Jul 2024).

5. Handling Free Boundaries and Cross-Embodiment Adaptation

Certain instances of ManiFlow—especially in simulation or manipulation—address the challenge of evolving fluid or object boundaries and cross-embodiment transfer:

• Free Boundaries in Meshfree Simulation: By maintaining dual point clouds representing fluid and underlying geometry, and employing local tessellation and projection schemes, ManiFlow methods robustly handle droplet splitting, merging, and moving boundaries on arbitrary surfaces (Suchde, 2020).
• Cross-Domain and Embodiment Adaptation: In manipulation, the flow-based abstraction (object flow or 3D flow) enables seamless transfer of manipulation skills between humans and robots, or across different robot architectures, by focusing on object-centric physical trends rather than embodiment-specific actions (Xu et al., 21 Jul 2024, Zhi et al., 6 Jun 2025, He et al., 14 Feb 2025). This representation supports large-scale pretraining on heterogeneous demonstrations and enables robust sim-to-real transfer.

6. Empirical Findings, Robustness, and Scaling

Across simulation and real-world deployments, ManiFlow systems demonstrate:

• Consistent empirical improvements in manipulation task success, with up to 45.6% improvement in dexterous simulation tasks and up to 98.3% increase in challenging bimanual setups (Yan et al., 1 Sep 2025).
• Robustness to novel objects, backgrounds, distractors, and shifts in environmental conditions as a function of both representation (manifold-respecting flows) and training regime (consistency, optimal transport, and multi-modal fusion).
• Scalability: Performance scales with dataset size, maintaining high success rates with hundreds of demonstrations and leveraging rich multi-modal data without overfitting or degradation in generalization.
• Efficiency: Flow matching with consistency allows action inference in as few as 1–2 steps, lowering latency in real-time applications.

7. Theoretical and Mathematical Underpinnings

ManiFlow approaches are grounded in rigorous mathematical formulations, blending concepts from

• Differential geometry: Tangential operators for PDEs on manifolds, explicit use of SO(3) and Lie algebra for rotation groups, geodesic embedding losses with autoencoders.
• Optimal transport: Dynamic formulation (Benamou–Brenier) with Wasserstein metrics used both in regularization and theoretical analysis.
• Flow-based generative modeling: Explicit change-of-variables, invertible mappings, and density estimation on manifolds, combined with stochastic differential transport (e.g., Neural ODE/SDE).
• Learning objectives: Joint optimization of flow matching and consistency losses, explicit handling of volume change in push-forward measures, and robust regularization for multimodal data and task conditioning.

The central loss functions in flow matching and consistency training are: $$\mathcal{L}_\text{FM} = \mathbb{E}_{x_0, x_1, t}\left[ \| v_\theta(x_t, t, 0) - (x_1 - x_0) \|^2 \right]$$

$$\mathcal{L}_\text{CT} = \mathbb{E}_{t, \Delta t}\left[ \| v_\theta(x_t, t, \Delta t) - \tilde{v}_\text{target} \|^2 \right]$$

where $\tilde{v}_\text{target}$ is computed from integration along flow trajectories.

---

ManiFlow encapsulates a versatile and mathematically principled toolkit for modeling, inference, and control on manifolds, consistently demonstrating improved geometric fidelity, robustness, and data efficiency across simulation, vision, shape, trajectory, and manipulation applications.