Preemptive Spatiotemporal Trajectory Adjustment

• Preemptive spatiotemporal trajectory adjustment is defined as modifying future trajectories using predicted occupancy and safety constraints to avoid future conflicts.
• The methodology leverages explicit four-dimensional representations and parallel optimization to evaluate spatial and temporal feasibility jointly.
• Applications span highway merging, human–robot collaboration, and aerial robotics, employing proactive maneuvers like lane changes, speed shifts, and route rerouting.

Searching arXiv for the cited works and closely related terminology to ground the article in current papers. {"query":"\"Preemptive Spatiotemporal Trajectory Adjustment\" OR trajectory adjustment system spatiotemporal", "max_results": 10} Searching for the main trajectory-planning and prediction papers referenced in the source corpus. {"query":"(Cao et al., 2019, Achaji et al., 2022, Zheng et al., 2023, Qian et al., 2019, Wu et al., 2024, Flowers et al., 2023, Arrizabalaga et al., 2022, Liang et al., 2021, Kuang et al., 5 Mar 2025)", "max_results": 20} A preemptive spatiotemporal trajectory adjustment system is a class of planning and control architectures that modifies future trajectories before conflicts, congestion, occlusions, or timing mismatches materialize. In the literature, the term appears explicitly in highway merging for heterogeneous autonomous vehicles, where vehicles and infrastructure coordinate space–time resources using safe spatiotemporal distances and type-aware cooperative control (Li et al., 30 Sep 2025). Closely related formulations appear in autonomous driving with discounted spatiotemporal safety costs and parallel nonlinear trajectory optimization (Zheng et al., 2023), in human–robot collaboration through time-indexed human occupancy avoidance (Flowers et al., 2023), in decentralized aerial robotics through spatiotemporal occupancy grid maps (Wu et al., 2024), and in active target tracking through visibility-aware planning on a four-dimensional spatiotemporal grid (Chen et al., 6 May 2026). This suggests that the concept is best understood as a systems family defined by anticipation, explicit time-indexed safety modeling, and receding or offline trajectory adjustment rather than by a single canonical algorithm.

1. Conceptual scope and defining characteristics

Across domains, the defining feature is not merely trajectory planning, but planning against a predicted future environment. In highway merging, the preemptive formulation coordinates ramp and mainline vehicles before they reach the confluence point, using a preset merge point, predicted arrival times, vehicle type, and a quantified safety spatiotemporal distance (Li et al., 30 Sep 2025). In human–robot collaboration, STAP computes avoidance intervals over workspace voxels and shifts edge traversal times to pass through predicted free time windows instead of reacting only to instantaneous human poses (Flowers et al., 2023). In congested autonomous driving, multiple candidate trajectories are optimized in parallel and then selected by safety, efficiency, comfort, lateral deviation, and consistency metrics, with the controller acting before barrier values collapse or TTC becomes critical (Zheng et al., 2023).

A second defining characteristic is explicit spatiotemporal representation. Some systems encode future occupancy directly as a four-dimensional grid indexed by space and time (Wu et al., 2024), while others encode future interactions as predicted trajectories, cluster sequences, or occupancy tubes (Cao et al., 2019). Others formulate the problem on state–time graphs, corridor–route topologies, or spatiotemporal lattices (Qian et al., 2019). The common structure is that feasibility is evaluated jointly over where an agent will be and when it will be there.

A third characteristic is anticipatory adjustment rather than purely reactive tracking. The adjustment may be a proactive lane change, a speed shift, a waiting action inserted before entering an occupied interval, a reroute across alternative homotopy classes, or a forecast-conditioned synthetic continuation of a mobility plan (Li et al., 30 Sep 2025). In this respect, preemption refers to acting upstream of a future incompatibility in space–time, not simply minimizing instantaneous tracking error.

2. System architecture and data representations

The data layer varies by application, but the underlying pipeline is structurally similar. GPS-based systems begin from timestamped point streams with device identifiers; TPM uses raw points with id, timestamp, and latitude/longitude, segments them into trajectories by a time-gap threshold $\theta$, filters noise, and then labels points with cluster IDs produced by K-means++ over spatial data (Cao et al., 2019). Scene-based motion prediction systems instead operate on synchronized multi-agent tensors: PreTR uses $$X \in \mathbb{R}^{T_{\text{obs}} \times N \times d_{\text{inp}}}$$, where each agent state contains 2D position and 2D velocity, while DA-STGCN models aircraft interactions through node features $v_t^i = [x_t^i, y_t^i, z_t^i]$ and time-indexed adjacency matrices derived from inverse-distance kernels (Achaji et al., 2022).

Time-indexed occupancy representations are central in safety-critical variants. In decentralized MAV planning, the environment is represented as a spatiotemporal occupancy grid map $\mathbb{M}_k$ with spatial resolution $r_s$, temporal resolution $r_\tau$, and a finite horizon of $T$ frames; communicated trajectories from other robots are projected into this map by Minkowski sums, yielding a unified occupancy representation for both obstacles and cooperative agents (Wu et al., 2024). In STAP, occupancy is represented not as a dense probability field but as voxel-wise avoidance intervals $A_{o,i}(x,y,z) = [t^s_{o,i}, t^f_{o,i}]$ and associated last-pass-times, which are then lifted to robot edges in configuration space (Flowers et al., 2023).

Road-network representations induce a different abstraction. NetTraj represents each vehicle trajectory as a sequence of intersections and associated movement directions, reducing the output space from city-scale road-segment vocabularies to a compact directional alphabet with $K=8$ directions (Liang et al., 2021). Geo-Llama goes further toward schedule-level modeling: a trajectory is a sequence of visits $v_j = (t_j, l_j, d_j)$, with location discretized into grid cells and duration as visit feature, enabling prompt-based insertion of explicit location–time constraints (Li et al., 2024).

These representations imply different forms of adjustment. Occupancy grids favor collision avoidance and short-horizon replanning; graph-based encodings favor route selection and interaction modeling; visit-level encodings favor schedule repair under explicit temporal constraints. This suggests that preemptive adjustment is representation-dependent: the same systems principle is instantiated differently depending on whether the operative object is a point trajectory, a path on a graph, a set of occupancy intervals, or a visit sequence.

3. Prediction and anticipation mechanisms

Prediction modules supply the future state hypotheses against which adjustment is performed. In crowded pedestrian forecasting, PreTR uses a factorized spatio-temporal Transformer with separate temporal and spatial self-attention blocks and a non-autoregressive decoder driven by learned object queries, thereby avoiding exposure bias. On ETH/UCY, it observes $$X \in \mathbb{R}^{T_{\text{obs}} \times N \times d_{\text{inp}}}$$0 steps and predicts $$X \in \mathbb{R}^{T_{\text{obs}} \times N \times d_{\text{inp}}}$$1 future steps, achieving average ADE/FDE of $$X \in \mathbb{R}^{T_{\text{obs}} \times N \times d_{\text{inp}}}$$2 m and reporting approximately $$X \in \mathbb{R}^{T_{\text{obs}} \times N \times d_{\text{inp}}}$$3 inference and per-batch training speedup over its autoregressive counterpart (Achaji et al., 2022). The model is deterministic and map-free, which is advantageous for low-latency anticipation but leaves multimodality and uncertainty calibration unmodeled.

In air-traffic prediction, DA-STGCN reconstructs the adjacency matrix through self-attention and combines STGCN with GAT and TXP-CNN to output probabilistic future aircraft positions. The future 3D state is modeled as a Gaussian with mean $$X \in \mathbb{R}^{T_{\text{obs}} \times N \times d_{\text{inp}}}$$4 and covariance $$X \in \mathbb{R}^{T_{\text{obs}} \times N \times d_{\text{inp}}}$$5, trained by negative log-likelihood over a $$X \in \mathbb{R}^{T_{\text{obs}} \times N \times d_{\text{inp}}}$$6-second horizon after a $$X \in \mathbb{R}^{T_{\text{obs}} \times N \times d_{\text{inp}}}$$7-second observation window (Kuang et al., 5 Mar 2025). The reported reductions are $$X \in \mathbb{R}^{T_{\text{obs}} \times N \times d_{\text{inp}}}$$8 in ADE and $$X \in \mathbb{R}^{T_{\text{obs}} \times N \times d_{\text{inp}}}$$9 in FDE relative to current 4D prediction methods, which is directly relevant because probabilistic forecasts can be translated into pairwise conflict probabilities or chance constraints.

Urban network prediction emphasizes topology rather than occupancy fields. NetTraj couples local graph attention over neighboring intersections with a sliding temporal attention window inside an LSTM encoder–decoder. On large-scale Shanghai and Beijing taxi datasets, it improves Distance Error and Average Match Ratio over road-network baselines, with especially strong gains in irregular networks (Liang et al., 2021). Because each predicted direction deterministically maps to a next intersection, the output can be used directly for route adjustment or congestion-aware navigation.

Not all systems begin with a predictive model. TPM mines dense urban regions, similar trajectories, and time-of-day distributions but does not include an explicit predictive module in its original form; its adaptation into a preemptive system requires added forecasting such as Markov chains, HMMs, or sequence models over cluster labels (Cao et al., 2019). This distinction matters: trajectory mining and trajectory adjustment are not equivalent. Preemptive systems require an anticipatory component, even if it is lightweight.

4. Safety modeling and trajectory adjustment mechanisms

The adjustment layer converts predictions into admissibility tests, costs, or optimization constraints. One recurrent pattern is explicit occupancy exclusion. In STAP, edge admissibility is defined by the absence of overlap between an edge traversal window $v_t^i = [x_t^i, y_t^i, z_t^i]$0 and the edge-level avoidance intervals $v_t^i = [x_t^i, y_t^i, z_t^i]$1, together with a last-pass-time condition $v_t^i = [x_t^i, y_t^i, z_t^i]$2. If overlap exists, the planner shifts the start time to the earliest feasible post-occupancy window and recomputes edge timing (Flowers et al., 2023). The same framework incorporates ISO/TS 15066 speed and separation monitoring through a maximum allowable tangential robot speed $v_t^i = [x_t^i, y_t^i, z_t^i]$3 that depends on separation distance, reaction time, deceleration capability, and human approach speed.

A second pattern is cost-based spatiotemporal safety. In dense autonomous driving, the safety module is built around an ellipse-based barrier function

$v_t^i = [x_t^i, y_t^i, z_t^i]$4

combined with time-discounted weights

$v_t^i = [x_t^i, y_t^i, z_t^i]$5

These produce running safety costs $v_t^i = [x_t^i, y_t^i, z_t^i]$6 inside a multiple-shooting nonlinear program solved in parallel for different lane hypotheses. The planner runs at $v_t^i = [x_t^i, y_t^i, z_t^i]$7 Hz with a $v_t^i = [x_t^i, y_t^i, z_t^i]$8 s horizon and $v_t^i = [x_t^i, y_t^i, z_t^i]$9 shooting steps, and reported simulations showed positive minimum barrier values, $\mathbb{M}_k$0 safety across tested algorithms, and higher safety margins for PTO6 than PTO1 (Zheng et al., 2023). The adjustment decision is then taken by multi-objective evaluation rather than by a single shortest-path criterion.

A third pattern is topology-aware maneuver restriction. SMSTP decomposes the environment into collision-free corridors in the $\mathbb{M}_k$1–$\mathbb{M}_k$2 plane and trajectory profiles in the $\mathbb{M}_k$3–$\mathbb{M}_k$4 plane, then groups these into maneuvers with explicit windows

$\mathbb{M}_k$5

Step-wise heuristic optimization first solves longitudinal motion and then lateral motion within these maneuver bounds, with the paper stating that a feasible trajectory is always guaranteed if a specific maneuver is given (Qian et al., 2019). This is a preemptive mechanism because topological infeasibility is removed before continuous optimization begins.

Occupancy-grid methods combine discrete and continuous layers. The decentralized MAV framework first runs kinodynamic A* against the spatiotemporal occupancy grid, then constructs safe flight corridors $\mathbb{M}_k$6, and finally solves a Bézier minimum-jerk quadratic program under corridor, velocity, and acceleration constraints (Wu et al., 2024). Because communicated robot trajectories are projected into the same map, robot–robot and robot–obstacle avoidance are unified at the representation level.

These mechanisms differ in mathematical form, but each operationalizes the same systems logic: predicted future occupancy or interaction structure is turned into hard exclusion, soft barriers, maneuver windows, or corridor constraints, and trajectory adjustment is solved before those predicted states are reached.

5. Representative domains and system realizations

The literature spans multiple operational domains rather than a single application family. The table summarizes representative formulations.

Domain	Representative mechanism	Reported outcome
Highway merging	Error-aware safe spatiotemporal distance, type-aware cooperative control	Maximum average delay improvement of $\mathbb{M}_k$7 for mainline and $\mathbb{M}_k$8 for ramp vehicles (Li et al., 30 Sep 2025)
Autonomous driving in congestion	Parallel NLP trajectory generation with discounted spatiotemporal safety barriers	Positive minimum barrier values and $\mathbb{M}_k$9 safety across tested algorithms (Zheng et al., 2023)
Human–robot collaboration	Predicted human occupancy intervals and SSM-aware spatiotemporal RRT*	STAP-IPTP reduced trajectory time by $r_s$0 on average relative to other methods (Flowers et al., 2023)
Decentralized MAVs	SOGM + kinodynamic A* + corridor-constrained Bézier QP	Average planning time about $r_s$1 ms per cycle in mixed environments (Wu et al., 2024)
Active target tracking	Beam-pruned layered DAG on a 4D grid with visibility scoring	$r_s$2 mean planning-time reduction and $r_s$3 convergence under tested cap (Chen et al., 6 May 2026)

The highway-merging formulation is the clearest explicit use of the term. It defines the safe spatiotemporal distance as

$r_s$4

and the static minimum merge gap as

$r_s$5

where $r_s$6 is the ramp-vehicle length (Li et al., 30 Sep 2025). The system then chooses mainline cooperative vehicles and, depending on whether the interacting vehicles are CAVs or CATs, either requests an inner-lane lane change or coordinates longitudinal acceleration and deceleration.

Human–robot collaboration uses the same preemptive idea in a different geometry. STAP plans through predicted free volumes rather than reserving highway gaps, and its real advantage is that it avoids both myopic reactive planning and the extreme conservatism of treating the union of all predicted occupancies as a static obstacle (Flowers et al., 2023). A plausible implication is that preemptive systems are most effective when future occupancy is neither ignored nor collapsed into a time-independent obstacle set.

Urban mobility and synthetic trajectory generation supply another variation. Geo-Llama formalizes controlled generation under a set of explicit visit constraints $r_s$7 and generates trajectories satisfying

$r_s$8

with prompts injecting visits at specified locations and times (Li et al., 2024). Although this is generative rather than control-oriented, it can function as a trajectory-adjustment backend when plans must be revised under dynamic time windows or exclusion zones. TPM, by contrast, offers dense-region mining and similar-trajectory extraction for navigation, recommendation, and urban resource allocation, but becomes preemptive only when paired with forecasting and triggering logic (Cao et al., 2019).

6. Evaluation practice, limitations, and open issues

Evaluation protocols vary sharply by domain, which complicates direct comparison. Prediction papers typically report displacement errors: PreTR uses ADE and FDE on ETH/UCY (Achaji et al., 2022), while DA-STGCN evaluates ADE/FDE separately for horizontal and vertical dimensions in ADS-B data (Kuang et al., 5 Mar 2025). Planning papers instead emphasize latency, convergence, safety margins, and task completion. The decentralized MAV framework reports success rates and average planning time around $r_s$9 ms (Wu et al., 2024); TA* reports $r_\tau$0 s wall time for $r_\tau$1 scenarios with $r_\tau$2 workers, $r_\tau$3 mean speedup over an unoptimized A* baseline, and only a $r_\tau$4 percentage-point average visibility change on the baseline-converged subset (Chen et al., 6 May 2026). Highway merging evaluates average speed and average delay, and STAP emphasizes actual-versus-estimated duration accuracy and average human–robot separation (Li et al., 30 Sep 2025).

Several limitations recur. Deterministic predictors remain common: PreTR outputs a single trajectory per agent and does not provide multimodal prediction or uncertainty calibration (Achaji et al., 2022). Simplified motion models are also widespread: the congested-driving framework uses constant-velocity prediction for surrounding vehicles (Zheng et al., 2023), and the MAV planner assumes constant-velocity prediction for non-cooperative dynamic obstacles (Wu et al., 2024). Such simplifications can be effective at short horizons but are brittle under abrupt intent changes.

Scalability and environment dependence are likewise unresolved. PreTR is trained with at most $r_\tau$5 agents per scene (Achaji et al., 2022); SOGM-based decentralized planning can fail by deadlock in dense clutter (Wu et al., 2024); TA* reports failure modes in Town07 dense vegetation, where visibility can collapse to zero under the given altitude and distance configuration (Chen et al., 6 May 2026). SMSTP states that its formulation cannot be directly used in urban traffic without HD-map topology because of complex intersections and sharp turns (Qian et al., 2019). Geo-Llama, although strong under explicit visit constraints, does not natively support hard exclusion constraints or sub-visit micro-mobility (Li et al., 2024).

A further misconception is that all preemptive systems are inherently predictive, robust, and privacy-preserving. The source literature does not support that view. TPM’s original formulation is decision-support-oriented rather than predictive (Cao et al., 2019). Geo-Llama is motivated partly by the inaccessibility and sensitivity of real mobility data, which underscores privacy concerns rather than resolving them automatically (Li et al., 2024). The broader picture is therefore heterogeneous: preemptive spatiotemporal adjustment is an organizing systems principle, but its concrete performance depends on the fidelity of prediction, the structure of safety constraints, the quality of map or occupancy representations, and the computational budget available for online replanning.

The current research trajectory points toward tighter coupling of probabilistic forecasting, explicit uncertainty handling, and dynamic optimization. That direction is already visible in the combination of probabilistic 4D aircraft prediction (Kuang et al., 5 Mar 2025), occupancy-based corridor planning (Wu et al., 2024), and parallel multi-hypothesis trajectory optimization (Zheng et al., 2023). This suggests that future systems will be defined less by a single planner and more by how effectively they integrate forecasting, safety envelopes, coordination, and receding-horizon adjustment into a unified space–time decision process.