TrajEvo

Evolving SOTA Trajectory Prediction Heuristics with LLMs

| Approach | Explainability | Speed | Performance | | :--- | :---: | :---: | :---: | | Manual Heuristic Design | ✅ | ✅ | ❌ | | Deep Learning | ❌ | ❌ | ✅ | | TrajEvo | ✅ | ✅ | ✅ |

TrajEvo takes the best of both worlds: it is fast and explainable like manual heuristic design, but also performs well like deep learning.

(more discussions in the paper)

Install

With uv:

bash uv venv --python 3.12 source .venv/bin/activate uv sync --all-extras

You may check the available LLM clients in utils/llm_client. Remember to set the environment variables for the LLM client you want to use. For example, to run the exp with VertexAI, remember to run export GOOGLE_APPLICATION_CREDENTIALS="..." with the json file.

Usage

From the root directory, you may run:

bash python main.py problem=trajectory_prediction init_pop_size=8 max_fe=50 timeout=20 llm_client=vertexai problem.dataset=eth Change problem.dataset to eth, hotel, univ, or zara1 for different datasets. This means training on all datasets except ETH (i.e., it will generate the "ETH" results).

Note that the main problem files repo for trajectory prediction is here: trajectory_prediction

Train TrajEvo (run multiple experiments)

bash python main.py -m problem=trajectory_prediction init_pop_size=8 max_fe=50 timeout=20 llm_client=vertexai problem.dataset=eth,zara1,zara2,hotel,univ seed=0,1,2 wandb_name="trajevo_v1"

The above will run multiple (with -m) experiments, across all datasets, with different seeds.

Evaluate TrajEvo

ETH-UCY

```bash

ETH

python trajectoryprediction/eval.py --codepath trajectoryprediction/trajevo/eth.py --codefunction predict_trajectory --test --dataset eth

Hotel

python trajectoryprediction/eval.py --codepath trajectoryprediction/trajevo/hotel.py --codefunction predict_trajectory --test --dataset hotel

Univ

python trajectoryprediction/eval.py --codepath trajectoryprediction/trajevo/univ.py --codefunction predict_trajectory --test --dataset univ

Zara1

python trajectoryprediction/eval.py --codepath trajectoryprediction/trajevo/zara1.py --codefunction predict_trajectory --test --dataset zara1

Zara2

python trajectoryprediction/eval.py --codepath trajectoryprediction/trajevo/zara2.py --codefunction predict_trajectory --test --dataset zara2 ```

This will evaluate TrajEvo held-out on each specific dataset.

SDD generalization

To evaluate the best heuristic found for the ETH split --> SDD generalization: bash python trajectory_prediction/eval.py --code_path trajectory_prediction/trajevo/eth.py --code_function predict_trajectory --test --dataset sdd

You can substitute the eth.py with other heuristics found in the trajectory_prediction/trajevo folder to obtain other results. Note that this specific value should be multiplied by 100 to convert to original scale.

Evaluate Heuristic Baselines

For CVM-S: bash python trajectory_prediction/eval.py --code_path trajectory_prediction/baselines/cvm/model.py --code_args "{'sample':True}" --test --code_function predict_trajectory --dataset eth

You may just change the code_path and the sampling attribute to obtain other baselines: - CVM: --code_path trajectory_prediction/baselines/cvm/model.py --code_args "{'sample':False}" - CVM-S: --code_path trajectory_prediction/baselines/cvm/model.py --code_args "{'sample':True}" - ConstantAcc: --code_path trajectory_prediction/baselines/constant_accel/model.py - CTRV: --code_path trajectory_prediction/baselines/ctrv/model.py - CSCRCTR: --code_path trajectory_prediction/baselines/cscrctr/model.py - SocialForce: --code_path trajectory_prediction/baselines/socialforce/model.py

As well as the dataset argument to evaluate on other datasets.

(Note: these results might be slightly different from what reported in the paper due to different random seed used at the time of writing)

Citation

If you find our work helpful (or if you are so kind as to offer us some encouragement), please consider giving us a star, and citing our paper:

bibtex @article{zhao2025trajevo, title={{TrajEvo}: Designing Trajectory Prediction Heuristics via LLM-driven Evolution}, author={Zhikai Zhao and Chuanbo Hua and Federico Berto and Kanghoon Lee and Zihan Ma and Jiachen Li and Jinkyoo Park}, journal={arXiv preprint arXiv:2505.04480}, url={https://github.com/ai4co/trajevo}, year={2025} }

Acknowledgements

Our work is built on top of the ReEvo repo! Please check it out: https://github.com/ai4co/reevo