Grafy News
Try Grafy free

Learned trajectory predictor beats kinematic models at 1 Hz V2X updates

A new arXiv study shows that a learned trajectory‑prediction model retains an advantage over constant‑velocity and constant‑acceleration baselines even when.

Illustration for the news story: Learned trajectory predictor beats kinematic models at 1 Hz V2X updatesSource: Grafy Research Lab

Why does observation frequency matter for autonomous‑driving prediction?

Trajectory prediction is a core component of automated driving stacks. Most models are trained on high‑frequency sensor logs—often 10 Hz or higher—because dense data makes it easier to learn motion patterns. In real‑world V2X (vehicle‑to‑everything) communication, however, position messages are typically sent at much lower rates, ranging from 1 Hz to 10 Hz. The gap between training conditions and deployment bandwidth can degrade prediction quality, potentially affecting safety and comfort. Understanding how reduced observation cadence impacts model performance is therefore critical for any system that relies on external motion cues.

How did the researchers evaluate the effect of lower sampling rates?

The study, authored by Nukhbala Gadzhiev and Alexander Morochko, used the Argoverse 2 motion‑forecasting dataset as a benchmark. They extracted a 5‑second history of the focal agent’s positions and then artificially down‑sampled this history from the original 10 Hz to three coarser rates: 5 Hz, 2 Hz, and 1 Hz. The task was to forecast the agent’s trajectory for the next 6 seconds. Crucially, the experiment omitted map information and neighboring agents, isolating the influence of the observation history itself. The authors released code and per‑scenario results alongside the paper, allowing others to reproduce the exact setup.

Which models were compared and why?

Three families of predictors were put side by side:

  1. Learned predictor – a data‑driven model trained on the full‑frequency dataset, which ingests the past positions and outputs future waypoints.
  2. Constant‑velocity (CV) model – a classic kinematic baseline that assumes the vehicle continues at its last observed speed and heading.
  3. Constant‑acceleration (CA) model – another simple physics‑based approach that extrapolates using the last observed acceleration.

The CV and CA models represent the traditional, computationally cheap methods often used when sensor data are scarce. By contrasting them with a modern learned approach, the authors could quantify whether sophisticated learning still offers a tangible benefit when only sparse observations are available.

What did the paper reveal about performance under reduced frequencies?

According to the authors, decreasing the observation frequency inevitably harms prediction accuracy for all three methods. Nevertheless, the learned predictor consistently outperformed both the CV and CA baselines across the 5 Hz, 2 Hz, and 1 Hz settings. The performance gap narrowed as the frequency dropped, indicating that the advantage of learning diminishes but does not disappear entirely. The results, presented in two tables and a single figure, suggest that even with V2X updates as low as 1 Hz, a model that has been trained on richer data can still provide more reliable forecasts than purely kinematic extrapolation.

How can developers and researchers experiment with these findings using Grafy?

Grafy offers several tools that make it easy to explore the paper’s methodology and reproduce its experiments:

  • Grafy Reader lets you upload the arXiv PDF, ask natural‑language questions about the experimental design, and retrieve specific sections without leaving the workspace.
  • Grafy Code provides an AI‑assisted coding environment where you can clone the authors’ repository, run the down‑sampling pipeline, and benchmark your own models against the CV and CA baselines.
  • For quick hypothesis testing, Grafy Chat can generate synthetic trajectory data at custom frequencies, helping you prototype variations of the experiment before committing to full training runs.

By leveraging these apps, you can verify the paper’s claims on your own datasets, extend the analysis to include map context or neighboring agents, and even integrate the learned predictor into a larger autonomous‑driving stack.

What are the practical takeaways for V2X‑enabled systems?

The study highlights three actionable insights:

  1. Don’t discard learned models – Even with sparse updates, data‑driven predictors remain superior to simple kinematic extrapolation.
  2. Design for frequency variability – Systems should be robust to a range of update rates; training on high‑frequency data and testing on down‑sampled versions can expose weaknesses early.
  3. Leverage external data when possible – While the experiment omitted map and neighbor information, incorporating these cues could further mitigate the loss caused by low‑frequency observations.

Developers building V2X‑based safety or navigation modules should thus consider retaining learned predictors and validate them across the full spectrum of realistic communication rates.


Source: Gadzhiev, N., & Morochko, A. “Learned versus Kinematic Trajectory Prediction under Reduced Observation Frequency,” arXiv:2611.23198 (2026).

Frequently asked questions

Does the learned predictor still work when V2X updates are only 1 Hz?

Yes. The paper reports that the learned model remains more accurate than constant‑velocity and constant‑acceleration baselines even at a 1 Hz observation rate, although the performance gap narrows compared to higher frequencies.

Why were maps and neighboring agents excluded from the experiment?

The authors wanted to isolate the effect of observation frequency on the predictor’s core ability to extrapolate motion from its own history. Removing external context eliminates confounding variables, making the comparison between learned and kinematic models clearer.

How can I test my own trajectory model with the same down‑sampling scheme?

Use Grafy Code to clone the authors’ repository, then apply the provided down‑sampling script to your dataset, selecting 10 Hz, 5 Hz, 2 Hz, or 1 Hz rates. The same evaluation pipeline can then measure error metrics against the CV and CA baselines.


Source: Grafy Research Lab — Learned versus Kinematic Trajectory Prediction under Reduced Observation Frequency