Dynamic M2M Velocity Fields

• Dynamic M2M velocity fields are time-indexed continuous representations that map local observations to global motion dynamics across space and time.
• They leverage probabilistic models, spline reconstructions, and neural networks to learn, infer, and optimize velocity and acceleration fields from scattered measurements.
• Utilized in robotics, galactic dynamics, and traffic analysis, they enable improved motion prediction, self-consistent modeling, and regime differentiation.

Dynamic M2M velocity fields are time-dependent velocity representations that map between motions, measurements, or model states through a field defined over space, time, or phase space. The expression is best treated as an Editor's term: in human-motion mapping it corresponds to a continuous probabilistic prior over velocities conditioned on location and time (Zhu et al., 16 Oct 2025); in particle-track reconstruction it denotes a measurement-to-model mapping from scattered observations to continuous velocity and acceleration fields (Gesemann, 2015); and in galactic dynamics it intersects with made-to-measure methods that recover self-consistent kinematic structure by adapting particle weights or masses (Hunt et al., 2012). Recent vision and robotics work extends the same core idea to 4D kinematic fields, Gaussian particle dynamics, and probability velocity fields learned from videos or cross-embodiment manipulation data (Im et al., 2024).

1. Terminology and conceptual scope

A central feature of the topic is terminological plurality rather than a single canonical formalism. In the spatio-temporal human-motion literature, the relevant object is a map that, given $(x,y,t)$, returns a full distribution over likely human velocities, including speed, heading, uncertainty, and multimodality; the cited work explicitly notes that the term “M2M” is not used there, but that the resulting representation is what one would want for a dynamic motion-to-motion velocity field (Zhu et al., 16 Oct 2025). In particle-based flow reconstruction, the same general idea appears as a weighted least-squares and penalized regression framework that maps measurements to model coefficients and can also be viewed symmetrically as model-to-measurement (Gesemann, 2015). In galactic dynamics, by contrast, M2M explicitly means made-to-measure: a live $N$-body model is adjusted so that observables at target particle locations are reproduced by a self-consistent dynamical system (Hunt et al., 2013).

A common misconception is that M2M names one method family. The literature instead uses the acronym in at least three ways: motion-to-motion in dynamic flow priors, measurement-to-model in reconstruction from observations, and made-to-measure in particle-based dynamical fitting (Zhu et al., 16 Oct 2025). What unifies these usages is not the acronym itself but the role of a velocity field as an intermediate object that couples local observations or states to global dynamics.

Across these traditions, dynamic M2M velocity fields are typically continuous or piecewise-continuous in time, support queries away from observed samples, and are used as priors, latent mechanisms, or self-consistency constraints rather than as mere post hoc derivatives of trajectories. This suggests that the most stable cross-domain definition is functional rather than terminological: a dynamic M2M velocity field is a time-indexed field whose evaluation mediates mappings among observations, configurations, or future states.

2. Mathematical representations

One major representation class is the continuous probabilistic Eulerian field. In the NeMo-map formulation, velocity is written in polar form as $\mathbf{v}=[\rho,\theta]^\top$, with speed $\rho$ and heading $\theta$, and the field is a neural map

$$\Phi_\theta : (\mathbf{x},t)\mapsto \{w_j(\mathbf{x},t),\mu_j(\mathbf{x},t),\Sigma_j(\mathbf{x},t)\}_{j=1}^J,$$

whose outputs parameterize a Semi-Wrapped Gaussian Mixture Model over velocities. The resulting conditional density

$$p(\mathbf{v}\mid \mathbf{x},t)=\sum_{j=1}^J w_j(\mathbf{x},t)\,N_{\mu_j(\mathbf{x},t),\Sigma_j(\mathbf{x},t)}(\mathbf{v})$$

defines a continuous, multimodal spatio-temporal velocity field over $(x,y,t)$ (Zhu et al., 16 Oct 2025). This representation is explicitly designed to capture opposing flows, sparse regions, and daily periodicity without discrete spatial cells.

A second class is the continuous deterministic or higher-order kinematic field. NVFi represents scene dynamics through a neural velocity field $\mathbf{v}(\mathbf{p},t)=g_\Phi(\mathbf{p},t)$, where $\mathbf{p}=(x,y,z)$, and couples it to a keyframe radiance field through advection and volume rendering (Li et al., 2023). Relatedly, a dynamic radiance-field model introduces a kinematic field

$N$0

so that velocity, acceleration, and jerk are primary outputs rather than finite-difference byproducts; displacements across arbitrary $N$1 are then approximated by a Taylor expansion in these quantities (Im et al., 2024). VeloGauss moves to a Lagrangian Gaussian-particle view and uses a physically grounded affine velocity model

$N$2

with translation, rotation, stretch, and shear basis functions, and per-particle Physics Codes that determine the time-varying coefficients (Lu et al., 11 May 2026).

A third class is spline- and kernel-based reconstruction. The trackfit/flowfit framework represents trajectories as temporal cubic B-splines and fields as spatial tensor-product B-splines,

$N$3

producing continuous-in-space velocity or acceleration fields from scattered measurements while retaining analytic derivatives (Gesemann, 2015). In particle-by-particle galactic M2M, the field is not written as a single closed-form Eulerian function; instead, local densities and velocity distributions are evaluated at target-particle positions with an SPH-type kernel, and the resulting kinematic structure is realized by a live $N$4-body system whose mass distribution and orbits co-evolve (Hunt et al., 2013).

A fourth class is the probability velocity field used in robot manipulation. There the basic object is a dense 2D image-plane velocity field over robot end-effector points, learned in a flow-matching formulation through the ODE

$N$5

The target field is locally stabilized around tracked trajectories by

$N$6

so the learned field is dense near the observed points and robust to small deviations (Seno et al., 22 Jun 2026).

3. Estimation, optimization, and inference

The estimation procedures used for dynamic M2M velocity fields are as heterogeneous as the representations themselves. Neural probabilistic flow maps are commonly learned by direct likelihood maximization. NeMo-map minimizes the negative log-likelihood of observed velocities under the predicted SWGMM,

$N$7

which directly fits the conditional velocity distribution rather than a single mean vector (Zhu et al., 16 Oct 2025).

Spline-based field reconstruction instead leads to sparse linear inverse problems. Trackfit estimates trajectory coefficients through a penalized least-squares functional with third-order finite-difference regularization, while flowfit solves a weighted least-squares problem for spatial B-spline coefficients and augments it with divergence-free, curl-free, and smoothness penalties. The normal equations remain sparse because B-splines have compact support and the penalty operators have local stencil structure (Gesemann, 2015).

In made-to-measure galactic dynamics, inference is embedded in the dynamics itself. Particle masses evolve under a force-of-change equation derived from density discrepancies, velocity likelihoods, temporal smoothing, and entropy regularization. The model therefore does not fit a static field in a fixed potential; it continuously adjusts a self-gravitating $N$8-body realization until local densities and velocity observables at target-particle locations are reproduced (Hunt et al., 2012). PRIMAL extends this by using likelihood-based velocity constraints in a barred disc and by aligning the model and target bar angles at every timestep during comparison (Hunt et al., 2013).

Video-based kinematic-field methods are trained from indirect supervision. NVFi couples a keyframe radiance field and an interframe velocity field through keyframe and interframe photometric losses, while also imposing divergence-free and momentum-conservation PINN terms on the learned velocity field (Li et al., 2023). The radiance-plus-kinematics formulation with $N$9 relies on photometric consistency, cycle consistency, and physics-driven regularizers for advective acceleration, jerk, transport, rigidity, and trajectory smoothness, without motion ground truth (Im et al., 2024). VeloGauss similarly learns geometry, appearance, and velocity from multi-view RGB videos, but enforces momentum transport and incompressibility through Global Physical Constraints evaluated at collocation points (Lu et al., 11 May 2026).

Robot-flow generation replaces reconstruction or likelihood fitting with conditional flow matching. The learned image-plane velocity field is trained to match an analytically constructed dense target field conditioned on initial image, goal image, and motion history, and a second flow-matching model maps the generated flow representation to robot actions (Seno et al., 22 Jun 2026). At a more statistical level, heterogeneous traffic fields have been modeled by combining an infinite HMM with multi-response Gaussian processes, with sequential MAP estimates for the hidden state sequence and efficient sequential GP posterior updates providing scalable inference (Chakraborty et al., 2021).

4. Temporal continuity, higher-order structure, and physical constraints

Temporal continuity is not an incidental detail; it is often the defining difference between a dynamic M2M velocity field and a time-binned map. NeMo-map is continuous in both space and time, using normalized time of day as an input coordinate and a SIREN temporal encoder to model daily periodicity. The stated motivation is precisely to avoid the discontinuities introduced by per-hour or per-bin maps such as CLiFF-map and STeF-map (Zhu et al., 16 Oct 2025).

Several frameworks go beyond instantaneous velocity and explicitly encode higher-order dynamics. The kinematic-field formulation for dynamic radiance fields treats velocity, acceleration, and jerk as continuous 4D fields and regularizes them with material-derivative identities,

$\mathbf{v}=[\rho,\theta]^\top$0

along with transport and rigidity terms derived from $\mathbf{v}=[\rho,\theta]^\top$1 and strain-rate invariants (Im et al., 2024). VeloGauss also constrains the field by continuum-mechanics-inspired equations, specifically momentum transport and incompressibility, but does so in a particle-centered Gaussian representation (Lu et al., 11 May 2026).

In reconstruction from experimental particle tracks, physical structure appears as soft penalties. Flowfit can enforce divergence-free velocity for incompressible flows and curl-free acceleration when acceleration is dominated by the pressure gradient, while still working with scattered measurements rather than gridded flow snapshots (Gesemann, 2015). A constructive variant appears in the spatially varying boost algorithm, which prescribes a bulk velocity field $\mathbf{v}=[\rho,\theta]^\top$2 by local Lorentz or Galilean boosts applied to otherwise non-moving field data; in the small-velocity regime this uses a scalar potential $\mathbf{v}=[\rho,\theta]^\top$3 with $\mathbf{v}=[\rho,\theta]^\top$4, and in the ultra-relativistic regime it is implemented as an infinitesimal-boost ODE in an auxiliary parameter (Ling, 28 Jun 2025).

Dynamic velocity fields also appear in theoretical PDE settings where the field is prescribed rather than inferred. The subsequentially fast dynamo construction produces a smooth, divergence-free $\mathbf{v}=[\rho,\theta]^\top$5 on $\mathbf{v}=[\rho,\theta]^\top$6 whose time-staged action on the induction equation yields exponential magnetic growth for suitable diffusivity sequences $\mathbf{v}=[\rho,\theta]^\top$7. There the field functions as a sequence of finite-time control maps on Fourier modes, rather than as a statistical prior or latent scene representation (Rowan, 29 May 2025).

5. Representative domains and reported behavior

Domain	Representative field form	Reported behavior
Human environments	Continuous SWGMM over $\mathbf{v}=[\rho,\theta]^\top$8	NLL $\mathbf{v}=[\rho,\theta]^\top$9
Robot manipulation	Image-plane probability velocity field	44 ms inference; 58% average success
Dynamic 3D scenes	4D velocity / kinematic / Gaussian particle fields	Future extrapolation and novel-view gains
Galactic discs	Particle-by-particle self-consistent kinematics	Recovery of radial profiles and bar pattern speed
Traffic	iHMM-GP regime fields	Interpretable regime switching on NGSIM

In human-environment modeling, the continuous probabilistic MoD learned on the ATC pedestrian dataset achieved $\rho$0 NLL, compared with $\rho$1 for Online CLiFF-map, $\rho$2 for CLiFF-map, and $\rho$3 for STeF-map, while also requiring about 19 minutes of training versus about 1831 minutes for batch CLiFF and approximately $\rho$4 seconds per query on GPU (Zhu et al., 16 Oct 2025). In this setting, the field acts as a site-specific spatio-temporal prior for robot navigation, collision avoidance, and long-horizon human motion prediction.

In robot manipulation, Flow as Flow treats robot motion as a dense probability velocity field in image space. On standard benchmarks it reports 44 ms inference, versus 1,430 ms for Track2Act, and in real-world experiments on 13 Toyota HSR tasks with 260 trials per method it attains a 58% average success rate, compared with 48% for Track2Act, 45% for original Im2Flow2Act, 28% for FLIP, 38% for DP-Lang, and 22% for DP-Goal (Seno et al., 22 Jun 2026). The stated significance is cross-embodiment transfer: the flow field is image-based and therefore embodiment-agnostic, while action generation remains embodiment-specific.

In dynamic scene understanding, explicit velocity fields support tasks beyond interpolation inside the training window. NVFi reports superior performance over all baselines, particularly in future frame extrapolation and unsupervised 3D semantic scene decomposition, because the velocity field is disentangled from the keyframe radiance field and reused for motion transfer and segmentation (Li et al., 2023). VeloGauss reports state-of-the-art performance in both Novel View Interpolation and Future Frame Extrapolation by combining Gaussian velocity fields, per-particle Physics Codes, and Global Physical Constraints (Lu et al., 11 May 2026).

In galactic dynamics, particle-by-particle made-to-measure methods recover global kinematic structure from discrete tracers. The initial particle-by-particle M2M method reconstructs the radial profiles of the surface density, radial and perpendicular velocity dispersions, and rotational velocity from a significantly mismatched initial disc (Hunt et al., 2012). PRIMAL extends this to barred galaxies and shows recovery of surface density, radial and vertical dispersions, mean rotational velocity, bar structure, and pattern speed, especially when the model bar is aligned to the target bar at every timestep (Hunt et al., 2013). With Gaia-like errors and extinction added to mock data, the modified PRIMAL still recovers these global properties to a reasonable degree of accuracy (Hunt et al., 2013).

In traffic analysis, the nonparametric Bayesian model treats each time slice as one latent velocity-field regime and learns both the regimes and their switching dynamics. Applied to NGSIM, it identifies interpretable field modes associated with free flow, turning, and transitional interaction patterns, and the learned transition structure reflects the temporal organization of multi-vehicle interactions (Chakraborty et al., 2021).

6. Limitations, misconceptions, and research directions

A recurrent misconception is that dynamic M2M velocity fields always encode explicit pairwise interaction laws. Several influential formulations instead encode interactions implicitly through shared fields, latent regimes, or global constraints. The traffic iHMM-GP model couples agents because they are noisy observations of the same latent field at each time, not because the model contains pairwise force terms (Chakraborty et al., 2021). VeloGauss likewise states that interactions are implicit and emergent through shared networks and Global Physical Constraints rather than through explicit pairwise forces (Lu et al., 11 May 2026).

Another misconception is that continuity guarantees adaptability. NeMo-map is trained in batch mode from long-term data and is not updated online, so site-specific priors can become outdated when environments change (Zhu et al., 16 Oct 2025). Flow as Flow, while efficient, is limited to 2D image-plane flows and assumes static cameras; strong camera motion, depth-sensitive tasks, and semantically under-specified goals remain failure modes (Seno et al., 22 Jun 2026). In galactic M2M, accuracy depends on assumptions about bar-angle alignment, known halo potentials, tracer populations, and observational errors; these are not incidental implementation details but structural conditions of the inference problem (Hunt et al., 2013).

Physical regularization is also domain-specific rather than universally transferable. Divergence-free velocity and curl-free acceleration are appropriate in the incompressible or pressure-gradient-dominated regimes targeted by flowfit, but not in arbitrary compressible dynamics (Gesemann, 2015). Similarly, rigidity penalties and incompressibility constraints improve scene reconstruction in many video settings, yet the same biases can oversimplify strongly deformable or highly chaotic motion (Im et al., 2024).

The main research directions are correspondingly diverse. In human-motion mapping, explicit online adaptation, richer context variables, and higher-order transition kernels have been proposed as extensions of continuous MoDs (Zhu et al., 16 Oct 2025). In spline-based reconstruction, full 4D spatio-temporal B-spline fields, boundary conditions, and additional Navier–Stokes-derived penalties are natural next steps (Gesemann, 2015). In dynamic Gaussian and radiance-field models, more explicit force laws, constitutive models, and larger-scale optimization remain open (Lu et al., 11 May 2026). In robot-flow models, multi-robot coordination, long-horizon flow chaining, and closed-loop online adaptation are direct extensions of the probability-velocity formulation (Seno et al., 22 Jun 2026). Taken together, these directions suggest that the field is moving from static or weakly regularized velocity maps toward structured, physically informed, and operationally integrated dynamic M2M velocity fields.