This repository contains the SF-LIFE dataset, a large-scale simulated movement dataset containing complete, noise-free trajectory data for 500,000 agents navigating the San Francisco Bay Area public transit network over a 70-day period. This README file contains an overview of the directory structure, and a brief look at the data structure in the different files.

For convenience, we provide several subsets of the data, both in terms of number of agents present and in terms of the sampling rates of the trajectories. These subsets are representative of the full data and should help reduce the data size according to the needs of the users.

Licenses and Attribution

SF-LIFE data

The SF-LIFE dataset, available in data/, is © 2026 The SF-LIFE authors and is made available under the Open Data Commons Attribution License (ODC-By): https://opendatacommons.org/licenses/by/ The license text is included in data/LICENSE_SF-LIFE_ODC-By.txt.

When using the SF-LIFE dataset, please cite the following paper or its follow-up journal version:

Chanuka Algama, Taylor Anderson, Henrique Ferraz de Arruda, Andrew Crooks, Nathan Holt, Erfan Hosseini Sereshgi, John Hunter, Hamdi Kavak, Lance Kennedy, Yueyang Liu, Dieter Pfoser, Sandro Martinelli Reia, Doug Taylor, Mauryan Uppalapati, Boyu Wang, Carola Wenk, and Andreas Züfle. SF-LIFE: A Large-Scale Simulated Movement Dataset for the San Francisco Bay Area. arXiv:2606.00430, 2026.

OpenStreetMap and OSM-derived data

The files in osm/ contain or are derived from OpenStreetMap data and are made available under the Open Database License (ODbL): https://www.openstreetmap.org/copyright The license text is included in osm/LICENSE_OSM_ODbL.txt.

In particular, osm/roads.osm contains road network data © OpenStreetMap contributors, and osm/sf-life_buildings.csv contains building centroids and categories derived by the SF-LIFE authors from OpenStreetMap geometries and metadata.

When using the files in osm/, please attribute OpenStreetMap contributors. When using the derived file osm/sf-life_buildings.csv please also attribute the SF-LIFE authors.

Directory Structure

README.md
osm
- LICENSE_OSM_ODbL.txt
- roads.osm
- sf-life_buildings.csv
data
- LICENSE_SF-LIFE_ODC_By.txt
- 500000_agents
  - sf-life_60_min_500000_agents_bucketed
    - agent_bucket_*.parquet
  - sf-life_30_min_500000_agents_bucketed
    - agent_bucket_*.parquet
  - sf-life_10_min_500000_agents_bucketed
    - agent_bucket_*.parquet
  - sf-life_1_min_500000_agents_bucketed
    - agent_bucket_*.parquet
  - sf-life_1_s_500000_agents_bucketed
    - agent_bucket_*.parquet
  - metadata
    - sf-life_agent_demographics_500000_agents.csv
    - sf-life_agenda_500000_agents.parquet
    - sf-life_agent_to_bucket_mapping_500000_agents.csv
- 10000_agents
  - sf-life_60min_10000_agents_bucketed
    - agent_bucket_*.parquet
  - sf-life_30min_10000_agents_bucketed
    - agent_bucket_*.parquet
  - sf-life_10min_10000_agents_bucketed
    - agent_bucket_*.parquet
  - sf-life_1min_10000_agents_bucketed
    - agent_bucket_*.parquet
  - sf-life_5s_10000_agents_bucketed
    - agent_bucket_*.parquet
  - metadata
    - sf-life_agent_demographics_10000_agents.csv
    - sf-life_agenda_10000_agents.parquet
    - sf-life_agent_to_bucket_mapping_10000_agents.csv
- 1000_agents
  - sf-life_60min_1000_agents_by_agent
    - agent_*.parquet
  - sf-life_30min_1000_agents_by_agent
    - agent_*.parquet
  - sf-life_10min_1000_agents_by_agent
    - agent_*.parquet
  - sf-life_1min_1000_agents_by_agent
    - agent_*.parquet
  - sf-life_5s_1000_agents_by_agent
    - agent_*.parquet
  - metadata
    - sf-life_agent_demographics_1000_agents.csv
    - sf-life_agenda_1000_agents.parquet
- 100_agents
  - sf-life_1min_100_agents_by_agent
    - agent_*.parquet
  - sf-life_5s_100_agents_by_agent
    - agent_*.parquet
  - metadata
    - sf-life_agent_demographics_100_agents.csv
    - sf-life_agenda_100_agents.parquet
- 15_agents
  - sf-life_1s_15_agents_by_agent
    - agent_*.parquet
  - metadata
    - sf-life_agent_demographics_15_agents.csv
    - sf-life_agenda_15_agents.parquet

Exploring the various files

In this section, we will look briefly at the various files present in the dataset, and explain the data found in each of them. We include sample Python code blocks for opening these files, and assume the same directory structure that is found in this repository.

All files can be loaded in Python using pandas:

import pandas as pd

Buildings file

This file defines the buildings used in the simulation. Agents may visit these buildings, and each building has a defined category such as residential, workplace, etc. We load and preview this file here:

buildings = pd.read_csv('osm/sf-life_buildings.csv')
print(buildings.shape)
print(buildings.head())
print(buildings.category.unique())

Output:

(159827, 4)
   buildingId   longitude   latitude     category
0           0 -122.471013  37.720996  residential
1           1 -122.468725  37.720688  residential
2           2 -122.468740  37.720784  residential
3           3 -122.468814  37.720605  residential
4           4 -122.468436  37.726570  residential
['residential' 'workplace' 'religion' 'education' 'restaurant' 'recreation']

The buildingId column acts as a primary key here which is referenced by other files in the dataset.

Demographics files

The demographics files define characteristics about each of the agents in the simulation. For subsets with fewer than the full 500,000 agents, it also defines which agents are present in that subset. We load one such file here, and one may load the demographics file for any given subset similarly:

demo_file = 'data/500000_agents/metadata/sf-life_agent_demographics_500000_agents.csv'
demographics = pd.read_csv(demo_file)
print(demographics.shape)
print(demographics.head())
print(demographics.agent_type.unique())

Output:

(500000, 5)
 agentId  age  gender  home_building agent_type
0      0    9       0         147152    student
1      1   32       0         147152     worker
2      2   15       1         147031     worker
3      3   32       1         147031     worker
4      4   33       0         147031     worker
['student' 'worker' 'homemaker']

The agentId column acts as a primary key here. We also have the age and gender for each agent (0 signifies male and 1 female), the home_building of their home location (which references the buildingId in the buildings file), and the agent_type, which is either student, worker, or homemaker. This determines typical behavior for an agent as well as the kinds of locations they may visit.

Bucket mapping files

For the subsets with 500,000 agents and 10,000 agents, we provide bucket files containing trajectories for several agents rather than a single trajectory file per agent, like we do with subsets containing fewer agents. For this reason, we also provide bucket mapping files for these subsets, to find in which bucket one may find the trajectory of an agent with a particular id. We load one such file here:

bucket_mapping = pd.read_csv('data/10000_agents/metadata/sf-life_agent_to_bucket_mapping_10000_agents.csv')
print(bucket_mapping.shape)
print(bucket_mapping.head())

Output:

(10000, 2)
   agentId  bucketId
0    85034        34
1    63572        22
2   164467        17
3   178440        40
4   130814        14

Agenda files

We provide some agenda files for this dataset. These files define the agenda which was used to simulate the different agents. For example, the agenda of a particular agent may say that the agent should be at home from midnight to 8 AM on a particular day, then spend 20 minutes walking to work, then stay at work from 8:20 AM to noon. The agent will try to fulfill this agenda during simulation, but will rarely fulfill these objectives with the exact timestamps provided. In some cases, agents will fail to reach some of these locations at all, for example if the travel time is longer than expected in the agenda and the agent needs to move to a new location before they arrive at their original intended destination.

We load and preview one such file here:

agenda_file = 'data/500000_agents/metadata/sf-life_agenda_500000_agents.parquet'
agenda = pd.read_parquet(agenda_file)
print(agenda.shape)
print(agenda.head())
print(agenda.activity_type.unique())
print(agenda.transport_mode.unique())

Output:

(152136152, 8)
   timestamp  agent  building activity_type transport_mode
0 2024-08-12      0    147152        AtHome           none 
1 2024-08-12      1    147152        AtHome           none
2 2024-08-12      2    147031        AtHome           none 
3 2024-08-12      3    147031        AtHome           none 
4 2024-08-12      4    147031        AtHome           none 
['AtHome' 'Transport' 'AtWorship' 'AtPudos' 'AtSchool' 'AtWork' 'AtErrand'
 'AtRestaurant' 'AtRecreation']
['none' 'multimodal' 'bicycle' 'auto' 'pedestrian']

We have several interesting columns here- first we have an agent and timestamp. This timestamp is the intended time for the agent to arrive at a location or start a trip. We then have a building column which references the buildings file. If the record in the file represents user movement, then the building column references the origin location.

The activity_type defines what the agent should be doing at the specified time. For instance, an agent may be AtHome, AtWork, or AtRestaurant. If the activity_type is Transport, then the agent should be moving between locations at this time. If the activity_type is AtPudos, the agent is either at a pick-up or drop-off for another agent.

The transport_mode defines how the agent should move between locations when activity_type is Transport, and should be 'none' otherwise. Typical cases here include bicycle, auto, and pedestrian, but if an agent should use multiple transportation modes on a trip, then transport_mode will be multimodal. Notably, an agent can also use public transport such as bus or rail, but in that case, they will always have to use another transport mode to reach their initial station and leave their final station, so this is always multimodal in the agenda.

Trajectory files

Finally, we observe the trajectories themselves. We load one example trajectory here:

traj = pd.read_parquet('/data/15_agents/sf-life_1s_15_agents_by_agent/agent_114451.parquet')
print(traj.head())

Output:

                  timestamp   agent  modality   longitude   latitude
0 2024-08-12 07:00:00+00:00  114451         0 -122.408528  37.802614
1 2024-08-12 07:00:01+00:00  114451         0 -122.408528  37.802614
2 2024-08-12 07:00:02+00:00  114451         0 -122.408528  37.802614
3 2024-08-12 07:00:03+00:00  114451         0 -122.408528  37.802614
4 2024-08-12 07:00:04+00:00  114451         0 -122.408528  37.802614

The trajecories have an agent and timestamp, as well as a longitude and latitude which define exactly where a particular agent is found at a particular time in the simulation. Finally, we have a modality, which is a number 0-6 and defines the agent's current transportation modality as follows:

0 = stationary
1 = walking
2 = bike
3 = car
5 = bus
6 = rail

Downloads last month: 164

Total file size:

499 GB

Paper for sf-life/sf-life

SF-LIFE: A Large-Scale Simulated Movement Dataset for the San Francisco Bay Area

Paper • 2606.00430 • Published 9 days ago