Motion Trajectory Field

• Motion Trajectory Field is a representation that encodes motion as continuous fields defined over state, space, time, or latent variables to capture dynamic behavior.
• It unifies various formulations—including vector fields, potential fields, and low-rank decompositions—to support applications in robotics, video processing, and human motion synthesis.
• This approach transforms discrete trajectories into spatially or temporally indexed functions, enabling queryable and robust planning and control.

Motion trajectory field denotes a family of motion representations in which trajectories are encoded as field-like objects over state, space, time, or latent variables. In the cited literature, the term covers several distinct but related constructions: a learned vector field / dynamical system for motion planning that drives a robot toward a demonstrated path and then to its endpoint; an object-conditioned implicit value function over $\mathrm{SE}(3)$ for reactive grasping; a scalar or vector field that unifies environmental, inertial, and social motion stimuli; a low-rank trajectory-space decomposition for dynamic 3D Gaussian primitives; and a dense mapping from every pixel in a video to a continuous 3D trajectory function of time (Li et al., 11 Sep 2025, Chen et al., 2022, Su et al., 2019, Li et al., 10 Aug 2025, Liu et al., 15 Oct 2025). Across these works, the common principle is that motion is represented as a queryable field rather than only as a discrete sequence of waypoints or framewise displacements.

1. Conceptual range of the term

Across recent work, “motion trajectory field” is not a single standardized object but a representation pattern. The query domain may be robot state, image pixels, 3D Gaussian anchors, gripper poses, or latent feature coordinates, while the field value may be a velocity update, a potential, a path-length cost, a continuous curve, or coefficients of a shared trajectory basis. This suggests a unifying view in which a trajectory field is any representation that turns motion into a spatially or temporally indexed function rather than a list of isolated samples.

Work	Domain	Field object
"KoopMotion: Learning Almost Divergence Free Koopman Flow Fields for Motion Planning" (Li et al., 11 Sep 2025)	Robot motion planning	Learned vector field / dynamical system
"Neural Motion Fields: Encoding Grasp Trajectories as Implicit Value Functions" (Chen et al., 2022)	Reactive grasping	Implicit value function over $\mathrm{SE}(3)$
"Potential Field: Interpretable and Unified Representation for Trajectory Prediction" (Su et al., 2019)	Trajectory prediction	Scalar potential field plus social force field
"3D Gaussian Representations with Motion Trajectory Field for Dynamic Scene Reconstruction" (Li et al., 10 Aug 2025)	Dynamic scene reconstruction	Shared trajectory bases with time-invariant coefficients
"Trace Anything: Representing Any Video in 4D via Trajectory Fields" (Liu et al., 15 Oct 2025)	4D video representation	Mapping from each pixel to a continuous 3D trajectory

A closely related strand treats motion as a mesoscopic vector field derived from trajectories. In "A generalized vector-field framework for mobility" (Liu et al., 2023), trajectories are converted into a local resultant field $W_i$ that captures both intensity and dominant local direction. In "A Tube-and-Droplet-based Approach for Representing and Analyzing Motion Trajectories" (Lin et al., 2016), a thermal transfer field encodes context-rich global motion patterns in a scene and is then used to derive a 3D tube and a droplet vector.

2. Dynamical-systems, potential-field, and contextual-field formulations

In robot planning, KoopMotion makes the “trajectory field” interpretation explicit. A demonstration is modeled as a discrete-time dynamical system,

$\mathbf{x}_{k+1} = f(\mathbf{x}_k),$

and nonlinear motion is parameterized through Koopman lifting,

$$\Psi(\mathbf{x}_{k+1}) = \mathcal{K}\Psi(\mathbf{x}_k).$$

The learned field is intended to attract states back to the demonstration manifold if they start off it, follow the demonstrated trajectory when already on it, and terminate at the desired goal state. The method adds a Koopman consistency loss, a goal fixed-point loss at $\mathbf{X}_{T_{\text{final}}}$, and an almost divergence-free regularizer,

$$\mathcal{L}_{FlowDivergence} = \Big\| \sum_i \frac{\delta \hat{F}_i}{\delta x_i} \Big\|_2,$$

to avoid local expansion, excessive contraction, and spurious attractors near the learned path (Li et al., 11 Sep 2025).

Potential-field formulations encode motion as a scalar landscape or force field from which future motion is decoded. "Potential Field" (Su et al., 2019) derives ground-truth potentials from observed trajectories, learns an environmental potential field from scene imagery, an inertial potential field from past motion, and a social force field

$\mathcal{F}\in\mathbb{R}^{H\times W\times 2}.$

Future motion is then obtained by predicting Gaussian distributions over direction and speed and rolling out

$$x_{\tau+1}=x_\tau+\mathcal{D}_\tau(x_\tau), \qquad \mathcal{D}_\tau=\mathcal{O}\cdot S(\tau)+\mathcal{F}.$$

The representation is explicitly presented as an interpretable, unified intermediate layer for environmental, inertial, and social stimuli.

Field-theoretic constructions also appear outside robotics. In urban mobility, local displacement vectors are aggregated within grid cells to produce

$$W_i = \frac{T_i}{\text{number of trips leaving } i},$$

and the resulting field is analyzed through divergence, curl, flux, and an approximate potential satisfying $\mathrm{SE}(3)$0 over most of the metropolitan space (Liu et al., 2023). In scene-level trajectory analysis, the thermal transfer field

$\mathrm{SE}(3)$1

encodes directional transfer coefficients derived from all trajectories in the scene; equipotential contours extracted from thermal diffusion maps are concatenated into a 3D tube, and the tube is compressed into a 36-dimensional droplet vector for clustering, classification, abnormality detection, and 3D action recognition (Lin et al., 2016).

3. Implicit value functions and learned motion fields for planning and control

A different interpretation replaces vector-field dynamics with a continuous cost field. "Neural Motion Fields" (Chen et al., 2022) defines an object-centric representation

$\mathrm{SE}(3)$2

where $\mathrm{SE}(3)$3 is a segmented object point cloud and $\mathrm{SE}(3)$4 is a gripper pose. The value of a pose is the path length of a trajectory from the current pose to a goal grasp pose, while a second branch predicts collision probability. The learned grasp cost is

$\mathrm{SE}(3)$5

with collision threshold $\mathrm{SE}(3)$6, and this cost is optimized inside a sampling-based MPC loop using STORM. Here the field is not a flow in state space but a continuous distribution over $\mathrm{SE}(3)$7 that can be queried online to choose the “best next pose.”

Flow-based generative planning generalizes the field concept to transport between trajectory distributions. "FlowMP: Learning Motion Fields for Robot Planning with Conditional Flow Matching" (Nguyen et al., 8 Mar 2025) represents joint-space trajectories with a spline basis and learns a continuous probability flow from a simple prior distribution $\mathrm{SE}(3)$8 to the expert trajectory distribution $\mathrm{SE}(3)$9. The motion field is decomposed as

$W_i$0

corresponding to velocity, acceleration, and jerk fields, and is trained by separate conditional flow-matching losses for first-, second-, and third-order dynamics. At inference, the learned field is integrated with an RK4-style scheme, and posterior guidance can be added through task costs. The paper explicitly contrasts this with first-order flow matching and diffusion-based planners, emphasizing smoother trajectories, smoother acceleration profiles, and more physically executable motions.

These two formulations share a continuous-query structure but differ in what is queried. Neural Motion Fields queries a gripper pose and returns a motion-to-grasp value and a collision estimate; FlowMP queries a transport state and returns higher-order motion increments. A plausible implication is that “trajectory field” in planning can denote either a feedback motion law or a continuous objective landscape, provided that the representation is dense enough to support replanning from unseen states.

4. Dense 4D video and dynamic-scene trajectory fields

In video representation, the term is made mathematically explicit. "Trace Anything" (Liu et al., 15 Oct 2025) defines a trajectory field

$W_i$1

so that every pixel in every frame is assigned a continuous 3D trajectory over normalized time. The curve is parameterized by spline control points,

$W_i$2

with cubic B-splines and clamped knots in the released model. This yields dense 3D trajectories, 3D point maps, 2D projections, scene flow, dynamic masks, and camera poses from a single feed-forward pass. The model is trained with a confidence-adjusted trajectory loss plus static, rigidity, correspondence, and timestamp regularizers, and the reported benchmark results include best video-mode $W_i$3, $W_i$4, $W_i$5, runtime $W_i$6 s, and image-pair runtime $W_i$7 s.

Dynamic scene reconstruction adopts a lower-rank variant of the same idea. "3D Gaussian Representations with Motion Trajectory Field for Dynamic Scene Reconstruction" (Li et al., 10 Aug 2025) expresses each point trajectory as

$W_i$8

and each Gaussian center as

$W_i$9

The coefficients $\mathbf{x}_{k+1} = f(\mathbf{x}_k),$0 are time-invariant and predicted from the Gaussian’s reference location, while the basis trajectories $\mathbf{x}_{k+1} = f(\mathbf{x}_k),$1 are shared across the scene. The same construction is extended to scale and rotation trajectories, static and dynamic Gaussians are separated using a probability $\mathbf{x}_{k+1} = f(\mathbf{x}_k),$2, and motion regularization combines segmentation supervision, an entropy-like regularizer, ARAP, and spatial smoothness. The reported ablation tied to static/dynamic separation shows PSNR improving from $\mathbf{x}_{k+1} = f(\mathbf{x}_k),$3 to $\mathbf{x}_{k+1} = f(\mathbf{x}_k),$4, and with $\mathbf{x}_{k+1} = f(\mathbf{x}_k),$5 reaching PSNR $\mathbf{x}_{k+1} = f(\mathbf{x}_k),$6, SSIM $\mathbf{x}_{k+1} = f(\mathbf{x}_k),$7, and FPS $\mathbf{x}_{k+1} = f(\mathbf{x}_k),$8.

These two works define opposite ends of a 4D spectrum. Trace Anything assigns a full continuous 3D curve to each pixel in each frame; the 3DGS trajectory field assigns a compact basis-coefficient trajectory model to each Gaussian anchor. In both cases, temporal structure is part of the representation itself rather than a post hoc association across frames.

5. Language, latent conditioning, and controllable generation

Trajectory fields have also become control interfaces for multimodal generation. "Lang2Motion: Bridging Language and Motion through Joint Embedding Spaces" (Galoaa et al., 11 Dec 2025) represents motion as dense tracked point trajectories $\mathbf{x}_{k+1} = f(\mathbf{x}_k),$9 with $$\Psi(\mathbf{x}_{k+1}) = \mathcal{K}\Psi(\mathbf{x}_k).$$0 points arranged in a $$\Psi(\mathbf{x}_{k+1}) = \mathcal{K}\Psi(\mathbf{x}_k).$$1 grid and $$\Psi(\mathbf{x}_{k+1}) = \mathcal{K}\Psi(\mathbf{x}_k).$$2 frames. A transformer-based auto-encoder maps trajectories into a CLIP-aligned latent space using both text supervision and rendered-trajectory image supervision, while the decoder predicts frame-to-frame displacements. The reported text-to-trajectory retrieval reaches $$\Psi(\mathbf{x}_{k+1}) = \mathcal{K}\Psi(\mathbf{x}_k).$$3, and generation quality includes $$\Psi(\mathbf{x}_{k+1}) = \mathcal{K}\Psi(\mathbf{x}_k).$$4, $$\Psi(\mathbf{x}_{k+1}) = \mathcal{K}\Psi(\mathbf{x}_k).$$5, $$\Psi(\mathbf{x}_{k+1}) = \mathcal{K}\Psi(\mathbf{x}_k).$$6, and $$\Psi(\mathbf{x}_{k+1}) = \mathcal{K}\Psi(\mathbf{x}_k).$$7.

"From Seeing to Predicting: A Vision-Language Framework for Trajectory Forecasting and Controlled Video Generation" (Yang et al., 1 Oct 2025) uses a Vision LLM built from SigLIP2 and Qwen2.5-8B to predict coarse-grained bounding-box trajectories

$$\Psi(\mathbf{x}_{k+1}) = \mathcal{K}\Psi(\mathbf{x}_k).$$8

together with chain-of-thought style reasoning. Those trajectories are then converted into masks that guide OpenSora through attention-based conditioning. The reported FVD scores are $$\Psi(\mathbf{x}_{k+1}) = \mathcal{K}\Psi(\mathbf{x}_k).$$9 on UCF-101 and $\mathbf{X}_{T_{\text{final}}}$0 on MSR-VTT, and removing the attention mask degrades performance from $\mathbf{X}_{T_{\text{final}}}$1 on UCF-101 and $\mathbf{X}_{T_{\text{final}}}$2 on MSR-VTT.

"Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance" (Chu et al., 9 Dec 2025) defines motion with dense point trajectories, projects them from pixel space into VAE latent coordinates, and then copies the first-frame latent feature at the trajectory start to later latent positions along the track. The resulting spatiotemporal feature map is the updated latent condition for the image-to-video model. Training trajectories come from CoTracker on a dense $\mathbf{X}_{T_{\text{final}}}$3 grid, and the guidance-strategy ablation reports EPE $\mathbf{X}_{T_{\text{final}}}$4 for pixel replication, EPE $\mathbf{X}_{T_{\text{final}}}$5 for random track embedding, and EPE $\mathbf{X}_{T_{\text{final}}}$6 for latent feature replication; with dense tracks $\mathbf{X}_{T_{\text{final}}}$7, EPE drops to $\mathbf{X}_{T_{\text{final}}}$8.

Human motion generation often uses trajectory conditioning without an explicit continuous field in the geometric sense. "TLControl: Trajectory and Language Control for Human Motion Synthesis" (Wan et al., 2023) learns a part-based VQ-VAE, predicts coarse latent motion codes from text plus masked partial trajectories with a Masked Trajectory Transformer, and then performs latent-space L-BFGS optimization to minimize control-joint position error. The reported runtime is $\mathbf{X}_{T_{\text{final}}}$9 s/frame and $$\mathcal{L}_{FlowDivergence} = \Big\| \sum_i \frac{\delta \hat{F}_i}{\delta x_i} \Big\|_2,$$0 FPS, with average errors around $$\mathcal{L}_{FlowDivergence} = \Big\| \sum_i \frac{\delta \hat{F}_i}{\delta x_i} \Big\|_2,$$1–$$\mathcal{L}_{FlowDivergence} = \Big\| \sum_i \frac{\delta \hat{F}_i}{\delta x_i} \Big\|_2,$$2 cm for several single-joint controls on HumanML3D. "Coordinating Multiple Conditions for Trajectory-Controlled Human Motion Generation" (Cai et al., 13 May 2026) separates trajectory control from motion completion: a first diffusion stage generates a simplified controlled-joint scaffold under trajectory guidance, and a second text-conditioned diffusion inpainting stage completes the body while freezing the observed controlled components. On HumanML3D, the reported pelvis-control results are FID $$\mathcal{L}_{FlowDivergence} = \Big\| \sum_i \frac{\delta \hat{F}_i}{\delta x_i} \Big\|_2,$$3, R-precision $$\mathcal{L}_{FlowDivergence} = \Big\| \sum_i \frac{\delta \hat{F}_i}{\delta x_i} \Big\|_2,$$4, foot skating $$\mathcal{L}_{FlowDivergence} = \Big\| \sum_i \frac{\delta \hat{F}_i}{\delta x_i} \Big\|_2,$$5, trajectory error $$\mathcal{L}_{FlowDivergence} = \Big\| \sum_i \frac{\delta \hat{F}_i}{\delta x_i} \Big\|_2,$$6, location error $$\mathcal{L}_{FlowDivergence} = \Big\| \sum_i \frac{\delta \hat{F}_i}{\delta x_i} \Big\|_2,$$7, and average error $$\mathcal{L}_{FlowDivergence} = \Big\| \sum_i \frac{\delta \hat{F}_i}{\delta x_i} \Big\|_2,$$8 cm.

Trajectory editing in video introduces yet another interpretation. "TrajectoryMover: Generative Movement of Object Trajectories in Videos" (Chhatre et al., 31 Mar 2026) treats editing as trajectory relocation: if the source object follows $$\mathcal{L}_{FlowDivergence} = \Big\| \sum_i \frac{\delta \hat{F}_i}{\delta x_i} \Big\|_2,$$9, the target is ideally

$\mathcal{F}\in\mathbb{R}^{H\times W\times 2}.$0

with adaptation when the shifted trajectory would create implausible interactions. The user interface is a source box and a target box in the first frame, and the model is trained on synthetic paired videos from TrajectoryAtlas. The reported metrics are $\mathcal{F}\in\mathbb{R}^{H\times W\times 2}.$1, $\mathcal{F}\in\mathbb{R}^{H\times W\times 2}.$2, $\mathcal{F}\in\mathbb{R}^{H\times W\times 2}.$3, and Bradley–Terry user-study utility $\mathcal{F}\in\mathbb{R}^{H\times W\times 2}.$4.

6. Evaluation criteria, misconceptions, and boundaries of the concept

The evaluation of motion trajectory fields is task-dependent because the represented object itself varies. In motion planning, KoopMotion is compared with DMP, CLF-DM, and CDSP using Dynamic Time Warping Distance and Swept Error Area; the central result is that DTWD is comparable to baselines while SEA is significantly better, indicating stronger spatiotemporal fidelity rather than only geometric curve matching (Li et al., 11 Sep 2025). In dense 4D video, Trace Anything uses $\mathcal{F}\in\mathbb{R}^{H\times W\times 2}.$5, $\mathcal{F}\in\mathbb{R}^{H\times W\times 2}.$6, $\mathcal{F}\in\mathbb{R}^{H\times W\times 2}.$7, Correspondence Agreement, and Static Degeneracy Deviation (Liu et al., 15 Oct 2025). In language-grounded generation, Lang2Motion uses Recall@K, ADE, FDE, Smoothness, CLIP similarity, AJ, and OA (Galoaa et al., 11 Dec 2025). In dynamic reconstruction, the 3DGS motion trajectory field is judged by PSNR, SSIM, LPIPS, and trajectory visualizations (Li et al., 10 Aug 2025). In human motion control, TLControl and CMC use trajectory error, location error, average error, FID, R-precision, Diversity, and foot skating ratio (Wan et al., 2023, Cai et al., 13 May 2026).

A recurrent misconception is that every trajectory-conditioned model is therefore a trajectory field. Several papers explicitly distinguish themselves from that interpretation. TLControl states that its notion of a “trajectory field” is not an explicit learned vector field in the usual sense and is closer to constrained latent optimization than to a classical trajectory field (Wan et al., 2023). CMC likewise does not explicitly use the terms “trajectory field” or “motion trajectory field,” and is best interpreted as a trajectory-conditioned control pipeline with an intermediate scaffold rather than a learned continuous field (Cai et al., 13 May 2026). "Analogical Trajectory Transfer" (Kim et al., 14 May 2026) also does not propose a motion trajectory field in the sense of a continuous field that directly stores or predicts trajectories everywhere in space; its closest field-like object is a smooth global warp $\mathcal{F}\in\mathbb{R}^{H\times W\times 2}.$8, assembled from per-cluster TPS maps, with runtime around $\mathcal{F}\in\mathbb{R}^{H\times W\times 2}.$9–$$x_{\tau+1}=x_\tau+\mathcal{D}_\tau(x_\tau), \qquad \mathcal{D}_\tau=\mathcal{O}\cdot S(\tau)+\mathcal{F}.$$0 seconds. TrajectoryMover similarly does not define a standalone explicit motion trajectory field, but instead learns a scene-aware latent motion transformation from source trajectory to relocated target trajectory (Chhatre et al., 31 Mar 2026).

The literature therefore supports a narrower and a broader reading of the term. In the narrow reading, a motion trajectory field is an explicit field-valued representation such as a learned vector field, a potential field, a dense pixel-to-curve map, or a low-rank trajectory-space decomposition. In the broader reading, it includes latent guides, value landscapes, and structured trajectory-conditioned intermediates that function like fields operationally even when they are not formulated as continuous fields mathematically. The cited work collectively indicates that the field abstraction is valuable precisely because it makes motion queryable away from demonstration points, away from observed frames, or away from a single prescribed path.