How Sparse V2X Data Tests Learned Vehicle Motion Prediction Models

Nukhbala Gadzhiev, Alexander Morochko

8 min read•Nov 18, 2026

A learned trajectory predictor retained a 24–30% advantage over the stronger kinematic baseline on straight and turning motion, even when position updates fell to 1 Hz. But prediction error still rose as updates became less frequent, and for pedestrians a simpler constant-velocity model was more accurate—showing that both update rate and the type of moving agent matter when choosing a predictor.

What did the researchers build?

The researchers tested how well different methods predict a moving agent’s future path when position updates arrive less often than the data used to train the predictor. This is a practical issue for vehicle-to-everything (V2X) communication, where reported positions can arrive at roughly 1–10 updates per second, while trajectory models are often trained on more frequent data.

The experiments used the Argoverse 2 motion forecasting dataset. For each scenario, the model received five seconds of position history for one focal agent and predicted the next six seconds. The researchers reduced the history from its original 10 Hz sampling rate to 5, 2, or 1 Hz. The prediction task used neither map information nor the positions of nearby agents, isolating how the focal agent’s own observed motion affected the forecast.

The learned method was an LSTM-based recurrent neural network (RNN), a model that processes a sequence of observations to estimate what comes next. The researchers compared models trained specifically for each observation frequency with a 10 Hz-trained model applied to sparse observations after linear interpolation. They also compared the learned predictor with two simple motion rules: constant velocity (CV) and constant acceleration (CA).

How did prediction accuracy change?

The frequency-specific RNN became less accurate as its input history grew sparser. Its average displacement error (ADE)—the average distance between predicted and actual positions over the forecast—rose by 3.1% at 5 Hz, 10.1% at 2 Hz, and 20.3% at 1 Hz, compared with the 10 Hz setting.

Despite that increase, the learned predictor kept a substantial advantage over the better of the two kinematic baselines on straight and turning motion. The reported advantage was 24–30% at every tested frequency. That advantage did not disappear on straight paths, contrary to the researchers’ initial expectation. The results were not uniform across agent types, however: for pedestrians, constant velocity was more accurate than the learned model.

Observation frequencyFrequency-specific RNN ADE change vs. 10 Hz10 Hz-trained RNN with interpolation
10 HzReferenceReference setting
5 Hz+3.1%At most 1.3% less accurate than the frequency-specific model
2 Hz+10.1%At most 1.3% less accurate than the frequency-specific model
1 Hz+20.3%11.5% less accurate than the frequency-specific model

The evaluation covered 24,984 test scenarios. The main RNN experiments used ten training seeds each, and the researchers calculated bootstrap confidence intervals over scenarios and seeds. The abstract does not report the interval bounds or absolute ADE values, so the relative changes show how performance shifted without giving a distance-based error that can be directly interpreted for a specific deployment.

How does the frequency comparison work?

The experiment separates two questions that can get mixed together in real deployments: what happens when a model sees fewer updates, and whether a model trained for frequent updates can still work when updates are sparse.

For the first question, the researchers downsampled each five-second history from 10 Hz to 5, 2, or 1 Hz. In practical terms, the model received fewer recorded positions describing the same stretch of motion. They trained a separate LSTM-based RNN at each frequency, then asked it to predict six seconds of future movement.

For the second question, they took the RNN trained at 10 Hz and supplied sparse histories after filling the gaps through linear interpolation. Interpolation estimates intermediate positions along a straight line between recorded points; it creates a denser input sequence, but does not add new observations of the agent’s actual movement. This setup tests whether a model trained on frequent data can tolerate a lower-rate input without retraining.

The study compared the learned predictions against CV and CA rules. CV extends the agent’s current velocity forward; CA also accounts for acceleration. These simple baselines are useful because they require little computation and provide a clear reference for whether the learned model adds predictive value.

The researchers set their hypotheses and decision thresholds before training. This reduces the chance of choosing success criteria after seeing the results. They also used multiple training seeds and bootstrap confidence intervals, which helps account for variation across both model runs and test scenarios. The setup remains deliberately focused: it examines a single agent’s observed positions, rather than the full set of maps, nearby traffic, and other signals available to a deployed driving system.

Why do these results matter for robotics?

Robots and automated vehicles often have to act on information that arrives at uneven or limited rates. V2X messages may provide useful position updates, but a planner still needs to estimate where another agent will be while the next message is on its way. This study shows that reducing the update rate can increase trajectory error, especially when it falls to 1 Hz, even when the prediction model has been trained specifically for that frequency.

The results also caution against assuming that one predictor is best for every moving object. A learned model retained a marked advantage over simple kinematic rules on straight and turning motion, but constant velocity did better for pedestrians in these experiments. An operating system that handles vehicles and people may therefore need to select or combine prediction methods based on the type of agent and the quality of its observations.

For engineering teams, the comparison between frequency-specific training and interpolation offers a practical clue: retraining for every update rate may not be necessary when inputs arrive at 5 or 2 Hz, based on this benchmark. At 1 Hz, the larger accuracy gap suggests that simply filling in missing positions is less effective. These findings are relevant to broader mobile automation, including systems buyers can browse among warehouse robots on Robot Overflow, where dependable motion estimates also support safe navigation and planning.

What are the study’s limitations?

The results come from one dataset and one constrained prediction setup: five seconds of focal-agent history, a six-second forecast, and no maps or neighboring agents. Those choices make it easier to isolate the effect of observation frequency, but they do not show how the same models perform when traffic interactions, road geometry, or additional sensor inputs shape movement.

The abstract reports relative ADE changes rather than absolute errors, and it does not provide the numerical bootstrap confidence intervals. That makes it difficult to judge the practical size of the errors in meters or assess uncertainty around the reported percentages. The experiments also do not establish that these predictions improve safety or control in a deployed vehicle. Finally, the pedestrian result shows that the learned model’s advantage should not be generalized to every agent type.

What do readers often ask?

Does the learned model outperform both kinematic baselines in every case? No. It outperformed the stronger kinematic model on straight and turning motion, but constant velocity was more accurate for pedestrians.

What update rates did the researchers test? They compared histories sampled at 10, 5, 2, and 1 Hz.

Did the predictor use maps or nearby agents? No. It used only the focal agent’s position history, so the experiment isolated prediction from that history.

Can a 10 Hz-trained model handle sparse input without retraining? It was at most 1.3% less accurate than a frequency-specific model down to 2 Hz, but 11.5% less accurate at 1 Hz.

Conclusion

The study finds that learned trajectory prediction can remain more accurate than simple motion rules even when position updates are sparse, but the margin depends on observation frequency and agent type. It also shows that interpolating missing inputs is a closer substitute for frequency-specific training at 2 Hz than at 1 Hz.