NGSIM is the Next Generation Simulation dataset, a free to use dataset created by the US Department of Transportation (DOT) and described here.

NGSIM traffic history scenarios

These scenarios are located in the smarts repository.

To use these scenarios, you must first download the dataset in an appropriate format. SMARTS assumes that the data will be in tab-separated text files. Its import scripts were tested on and work with the data downloaded from here, which is the same as was used by this PPUU project.

After the dataset has been downloaded, you will need to update each in the NGSIM scenario directories (e.g., NGSIM/i80/, NGSIM/us101/, and/or NGSIM/peachtree/ to point the input_path field to the dataset location on your filesystem.

Once that is done, you should be able to build your scenarios in the normal way, for example: scl scenario build-all --clean scenarios/NGSIM

For each traffic history dataset specified in your, this will create a corresponding .shf file that SMARTS will use whenever this traffic will be added to a simulation.

Note that the SUMO maps created for NGSIM were made by hand by the open-source SMARTS team. Although they attempt to align with the positions in the traffic dataset, their level of exactness may not be enough for some model-training situations, so you may want or need to refine them with SUMO’s netedit tool.

An example of how traffic history might be saved as observations can be found in traffic_histories_to_observations.

To consume the generated observations you could use the following approach:

import os
import pickle
import re

# a suggested approach
import numpy as np
from PIL import Image

from smarts.core.observations import Observation, TopDownRGB

scenarios = list()
for scenario_name in os.listdir(input_path):

for scenario in scenarios:
    obs: List[Observation] = list()
    vehicle_ids = list()

    for filename in os.listdir(scenario_path):
        if filename.endswith(".pkl"):
            match ="(.*).pkl", filename)
            assert match is not None
            vehicle_id =
            if vehicle_id not in vehicle_ids:

    for id_ in vehicle_ids:
        with open(scenario_path / f"{id}.pkl", "rb") as f:
            vehicle_data = pickle.load(f)

        image_names = list()
        for filename in os.listdir(scenario_path):
            if filename.endswith(f"{id_}.png"):

        for i in range(len(image_names)):
            with / image_names[i], "r") as image:
                bev = np.moveaxis(np.asarray(image, dtype=np.uint8), -1, 0)

            sim_time = image_names[i].split("_")[0]
            current_obs: Observation = vehicle_data[float(sim_time)]
            obs.append((current_obs, bev))

    for o, rgb in obs:

Alternatively, an approach like direct/ can be used to operate directly with the scenarios.


Some specific dataset samples can be found at: