TrajEvo: Designing Trajectory Prediction Heuristics via LLM-driven Evolution

TL;DR

TrajEvo addresses the gap between handcrafted trajectory heuristics and data-driven models by automatically designing prediction heuristics with an LLM-guided evolutionary process. It introduces Cross-Generation Elite Sampling and a Statistics Feedback Loop to evolve diverse, interpretable heuristics directly from data. On ETH-UCY, TrajEvo outperforms traditional heuristics and shows strong cross-dataset generalization to the unseen SDD, often outperforming DL baselines while delivering fast, explainable predictions. This work demonstrates a viable path toward automated, scalable, and interpretable trajectory prediction for real-time robotics.

Abstract

Trajectory prediction is a crucial task in modeling human behavior, especially in fields as social robotics and autonomous vehicle navigation. Traditional heuristics based on handcrafted rules often lack accuracy, while recently proposed deep learning approaches suffer from computational cost, lack of explainability, and generalization issues that limit their practical adoption. In this paper, we introduce TrajEvo, a framework that leverages Large Language Models (LLMs) to automatically design trajectory prediction heuristics. TrajEvo employs an evolutionary algorithm to generate and refine prediction heuristics from past trajectory data. We introduce a Cross-Generation Elite Sampling to promote population diversity and a Statistics Feedback Loop allowing the LLM to analyze alternative predictions. Our evaluations show TrajEvo outperforms previous heuristic methods on the ETH-UCY datasets, and remarkably outperforms both heuristics and deep learning methods when generalizing to the unseen SDD dataset. TrajEvo represents a first step toward automated design of fast, explainable, and generalizable trajectory prediction heuristics. We make our source code publicly available to foster future research at https://github.com/ai4co/trajevo.

Figures (5)

• Figure 1: Motivation for TrajEvo. Traditional manual heuristic design (left) is based on human experts with trial and error. Deep learning (center) generates better predictions but requires significant computational resources, generates black-box models, and struggles with generalization. TrajEvo (right) automates the design process of heuristics via evolutionary algorithms, generating novel trajectory prediction heuristics.
• Figure 2: Example evolution of trajectory prediction heuristics with TrajEvo.
• Figure 3: Cross-Generation Elite Sampling (CGES) helps escape local optima by sampling elite individuals from past generations (left), which greatly helps achieve much better objective values (right).
• Figure 4: Statistics obtained by running TrajEvo-generated code are provided alongside the corresponding code as an input to the reflector (left). TrajEvo then analyses this and gathers insights into how to evolve better trajectory prediction heuristics (right).
• Figure 5: Comparison of trajectory prediction results between CVM-S scholler2020constant and TrajEvo across different datasets. Each row illustrates a distinct behavioral pattern: (top) linear trajectories, (middle) non-linear trajectories, and (bottom) collision-avoidance cases. For each method, we report the single best trajectory out of $K=20$ samples based on the objective value $J$.