← News

Agents & MCPSources: arXiv · 18 Nov 2026

GCERT transformer targets map-free trajectory prediction for autonomous driving

Researchers propose a Goal-Conditioned Expert-Routed Transformer that predicts vehicle trajectories without high-definition maps, recovering motion and interaction cues the maps normally provide.

In plain words

Self-driving systems often predict where nearby vehicles will move, and many models rely on detailed high-definition maps to do it. This paper describes a model that makes those predictions without such maps, learning road context from short motion histories instead. The work is described in an arXiv abstract, so results are not independently confirmed here.

What was announced

The authors frame map-free prediction as a trade-off: dropping high-definition maps makes deployment more flexible but removes road-topology and lane-connectivity priors, so the model must recover stable motion cues from short histories and resolve interactions without lane-level guidance. Their framework first decomposes each agent's historical motion into a smooth trend and a local detail residual using a centered moving average, then processes the recombined signals with a causal temporal Transformer.

The core Goal-Conditioned Expert-Routed Transformer uses dual-level expert routing, where edge-level relation experts generate interaction-specific messages and node-level state experts adapt transformations after aggregation. Predicted motion goals update node representations and build relative-geometry and orientation edge features before the final interaction layer. The model outputs multimodal future trajectories and is trained jointly on trajectory-regression, mode-classification, and goal-prediction objectives. The abstract reports the architecture and training design but no benchmark results, so real-world performance remains unverified.

What it does not show

  • The source is an abstract only; no experimental results, datasets, or comparisons are provided, so the framework's effectiveness cannot be assessed from this evidence.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.