Velocity-Field DVE for Autonomous Planning

• The paper introduces the DVE model that represents preferred velocities as an implicit function, encoding both heading and speed based on contextual data.
• It employs dual loss functions—velocity-imitation and velocity-correction—to learn context-dependent velocity fields effectively from demonstration trajectories.
• Empirical results on the Waymo dataset show significant improvements in collision avoidance and planning accuracy compared to traditional scalar cost-field approaches.

A velocity-field DVE (Dynamic Velocity Environment) is an implicit representation of preferred or recommended velocities—both magnitude and direction—through a spatial domain, often extended in time, for the purposes of trajectory planning, simulation, or analysis. The DVE concept has most recently been formalized in the context of autonomous vehicle planning as an implicit queryable field $\mathbf{v} = f(x, y, t|\mathcal{S})$, where $(x, y)$ are coordinates, $t$ is time, and $\mathcal{S}$ encodes global scene context. The DVE differs fundamentally from scalar cost fields or rasterized maps by encoding vector-valued priors at each query point, reflecting both heading and speed derived from demonstration or learned policy (Xin et al., 2024). This article presents a comprehensive survey of the DVE concept with a focus on velocity-field representations, their role in planning architectures, derivation from data, algorithmic integration, and empirical benchmarking.

1. Mathematical Definition and Formalization

A Dynamic Velocity Environment is a function

$\mathrm{DVE}: (x, y, t) \mapsto (u, v)$

parameterized by a latent context vector $\mathcal{S}$ derived from the map, agent history, and dynamic elements (e.g., other vehicles, traffic lights). The velocity field is implicitly modeled, typically by a neural network with cross-attention mechanisms, as

$$\mathbf{v}(x, y, t) = f\big([x, y, t, \sin t, \cos t] ~|~ \mathcal{S}\big),$$

outputting preferred longitudinal and lateral velocity components at any spatial-temporal coordinate. In its basic static form, $\mathbf{v}(x, y) = (u(x, y), v(x, y))$, but to capture realistic driving context and agent interaction it is extended to include temporal and scene dependencies (Xin et al., 2024).

2. Learning Velocity Fields from Demonstration

The DVE representation is trained using demonstration trajectories $\{\mathbf{p}_e^t\}_{t=0}^T$. For each time $t$ and sampled point $(x, y)$0 near the demonstration, the DVE field is queried to provide $(x, y)$1. Two losses are applied:

• Velocity-imitation loss: Encourages the field at and near the demonstration to approximate the demonstrator's velocity vector, weighted by spatial (or yaw) proximity:

$(x, y)$2

• Velocity-correction loss: Forces the field to point toward the next demonstrator position, again proximity-weighted:

$(x, y)$3

The total loss is $(x, y)$4. This methodology allows the DVE to encode emergent priors for both heading and speed directly from data, and the implicit neural parameterization sidesteps the grid-resolution tradeoffs of rasterized map approaches (Xin et al., 2024).

3. From Velocity Field to Cost and Planning

To utilize the DVE in trajectory planning, a trajectory $(x, y)$5 is evaluated by comparing its instantaneous velocity $(x, y)$6 with the local field value:

$(x, y)$7

A Lagrangian cost $(x, y)$8 accumulates this velocity-mismatch term along with classical smoothness regularizers (acceleration, jerk, steering angle change):

$(x, y)$9

The total cost for a trajectory is $t$0 with $t$1 learnable or hand-tuned (Xin et al., 2024).

A STOMP-style optimizer is used in control space (acceleration, steering), iteratively sampling a batch of trajectories, evaluating cost via DVE queries, and updating the distribution over controls toward lower-cost solutions. This process yields plans that are context-aware and human-like in heading and speed modulation.

4. Algorithmic Integration and Implementation

The velocity-field DVE is integrated into planning pipelines as an implicit, queryable prior for control or trajectory optimization. At each planning iteration:

• Candidate control sequences are sampled.
• Resulting trajectories are rolled out in simulation.
• At each trajectory point and time step, the DVE is queried for the preferred velocity vector in context.
• Costs are computed and top-performing controls selected for subsequent sampling.

This approach decouples the environment and motion representation from the cost-mapping paradigm, providing richer guidance especially in multimodal or dynamically constrained scenes. Unlike raster cost maps, the implicit DVE is memory efficient and scales gracefully with spatial-temporal domain size, as no explicit grid decoding is needed (Xin et al., 2024).

5. Empirical Results and Comparative Analysis

Evaluation on the Waymo Open Motion Dataset demonstrates that DVE-based planners significantly improve safety, comfort, and plan accuracy:

Method	Collision ↓	Off-route ↓	TL Violation ↓	ADE ↓	FDE ↓	L2@1s ↓	L2@3s ↓	L2@5s ↓	Avg. Runtime (s) ↓
VF (DVE)	8.01%	1.65%	1.41%	0.7071	1.7555	0.1371	0.8743	2.4642	0.048
CF	6.26%	0.32%	1.61%	0.7379	1.8579	0.4014	3.0564	8.2568	—

The DVE reduces collision rate by 33.3% and red-light violations by 80% compared to a scalar cost-field baseline, with the lowest planning error at all tested time horizons. Computational efficiency is maintained, with real-time planning feasible at 0.048 s average per call on contemporary hardware (Xin et al., 2024).

6. Distinctions from Related Concepts and Broader Applications

The term DVE as introduced in (Xin et al., 2024) denotes an implicit velocity field parameterized by environment context and learned from demonstration, distinguishing it from:

• Scalar cost fields (CF): Encode only obstacle or heuristic cost, not vectorial guidance; DVE encodes both direction and speed priors.
• Rasterized cost maps: Require explicit decoding and are subject to spatial resolution-memory tradeoffs; DVE as an implicit function allows arbitrary granularity without grid artifacts.
• Physics-based velocity fields: Appear in fluid or traffic modeling contexts; DVE, as defined here, specifically targets learning from expert trajectories under high-dimensional context for decision-making and planning.

While the DVE formulation is specifically motivated by autonomous driving and trajectory planning, the principle of learning context-dependent, implicit vector fields for guidance naturally generalizes to other domains involving navigation, multi-agent coordination, and continuous control.

7. Significance and Implications

The introduction of the velocity-field DVE addresses core limitations in cost-centric planning and trajectory optimization in high-dimensional state spaces. The DVE approach synthesizes scene, map, and dynamic agent information into a low-latency, queryable field, providing richer directional and speed priors than traditional methods. Empirical results confirm its effectiveness in safety-critical planning, with substantial improvements in collision avoidance, adherence to route, and traffic law compliance. This suggests that implicit dynamic velocity fields provide a scalable and semantically meaningful intermediate representation for modern planning and control architectures (Xin et al., 2024).