NeMo-map: Neural Implicit Flow Fields for Spatio-Temporal Motion Mapping

Yufei Zhu · Shih-Min Yang · Andrey Rudenko · Tomasz P. Kucner · Achim J. Lilienthal · Martin Magnusson

RNP@ORU

The official implementation of ICLR'26 paper: "NeMo-map: Neural Implicit Flow Fields for Spatio-Temporal Motion Mapping"

📝 Overview

Safe and efficient robot operation in complex human environments can benefit from good models of site-specific motion patterns. Maps of Dynamics (MoDs) provide such models by encoding statistical motion patterns in a map, but existing representations use discrete spatial sampling and typically require costly offline construction. We propose a continuous spatio-temporal MoD representation based on implicit neural functions that directly map coordinates to the parameters of a Semi-Wrapped Gaussian Mixture Model. This removes the need for discretization and imputation for unevenly sampled regions, enabling smooth generalization across both space and time. Evaluated on two public datasets with real-world people tracking data, our method achieves better accuracy of motion representation and smoother velocity distributions in sparse regions while still being computationally efficient, compared to available baselines. The proposed approach demonstrates a powerful and efficient way of modeling complex human motion patterns and high performance in the trajectory prediction downstream task.

🛠️ Run the experiment

Clone the code

after git clone, also do git lfs pull, to pull the large data files in atc/

Setup environment

conda env create -f environment.yml
conda activate mod

For training

python3 train.py  --model siren

The argument --model can be one of ["siren", "time_grid", "fourier"]. In the paper, we use siren as the main model. For the ablation study, we provide two alternative temporal encodings: time grid and Fourier features.

The training runs for 100 epochs, and models are saved in models/distri_gmm_siren. For time_grid and fourier version, trained models are also provided models/distri_gmm_feature_ff_time, and models/distri_gmm_feature_time.

For evaluation

python3 evaluate_NLL.py --model siren

After training the model, we can evaluate it by computing the Negative Log Likelihood (NLL) value. Same here, the arg can be chosen from ["siren", "time_grid", "fourier"]. Detailed NLL results for each test sample will be saved in nll_results/distri_gmm_siren/atc-all.csv.

For querying MoDs

python3 generate_MoD_files.py --model siren

We can query the trained model to generated maps of dynamics for each hour of the ATC dataset. The generated MoDs are saved in MoDs/distri_gmm_siren/<hour>.csv

For plotting

python3 plot_MoD_files.py --model siren --version max

We can also plot the generated MoDs. Two version of plotting are provided.

• Version all shows multimodality by rendering all SWGMM components with transparency proportional to their weights.
• Version max more clearly shows the dominant flow, only displaying the mixture component with the largest weight.

The generated MoD figures are saved in MoDs/distri_gmm_siren/all_png and MoDs/distri_gmm_siren/max_png folders.