arena-rosnav/PEDS37

PEDS-37

Pedestrian and robot trajectories from PEDS over 37 combined scenarios.

Stats

• Trials: 2850
• Scenarios: bt_dense_group_conversation, bt_dense_queue_use, bt_dense_robot_group_conversation, bt_dense_robot_queue_use, bt_dense_robot_service_mobile, bt_dense_service_mobile, bt_sparse_compound, bt_sparse_group_conversation, bt_sparse_needs, bt_sparse_pair, bt_sparse_queue_use, bt_sparse_robot_compound, bt_sparse_robot_group_conversation, bt_sparse_robot_needs, bt_sparse_robot_pair, bt_sparse_robot_queue_use, bt_sparse_robot_service_mobile, bt_sparse_robot_service_static, bt_sparse_robot_sit, bt_sparse_service_mobile, bt_sparse_service_static, bt_sparse_sit, het_dense_mixed_speeds, het_dense_robot_in_group, het_dense_wheelchair_bottleneck, nav_dense_bottleneck, nav_dense_corridor, nav_dense_crossing, nav_dense_flow, nav_dense_robot_bottleneck, nav_dense_robot_corridor, nav_dense_robot_crossing, nav_dense_robot_flow, nav_sparse_bend, nav_sparse_corridor, nav_sparse_crossing, nav_sparse_flow, nav_sparse_merge, nav_sparse_robot_bend, nav_sparse_robot_corridor, nav_sparse_robot_crossing, nav_sparse_robot_flow, nav_sparse_robot_merge
• Pedestrian planners: hsfm, nsp, orca, sfm, socialgail, straight
• Mode: robots (heterogeneous human+robot)

Layout

• data/agent_states/bucket=<b>/density=<d>/scenario=<s>/planner=<p>/*.parquet - long-format trajectories, ~100_000 rows per row-group, zstd compressed. Partition values for bucket, density, scenario and planner are encoded in the path; mirrors the config/evaluation/<bucket>/<density>/<modality>.yaml input tree.
• data/metrics/kinematics_per_trial.parquet - per-trial scalar metrics (jerk, curvature, collisions).
• data/metrics/robot_metrics.parquet - per-trial robot KPIs (success, time-to-goal, path efficiency, personal-space violations). Robots-mode only.
• data/metrics/failures.parquet - per-trial failure cause classification. Robots-mode only.
• configs/scenarios/*.yaml - scenario YAML snapshots used at sweep time.
• croissant.json - Croissant 1.0 machine-readable schema.
• DATASHEET.md - Datasheet for Datasets (Gebru et al.) covering motivation, composition, collection, uses, distribution, maintenance.

Schema

agent_states (long-format trajectories)

field	dtype	unit	description
time	float64	s	sim time, monotonic per trial
agent_id	int64	-	unique within trial; humans use positive ids (auto-assigned from 1), robots use negative ids
x	float64	m	world-frame x position
y	float64	m	world-frame y position
vx	float64	m/s	world-frame x velocity
vy	float64	m/s	world-frame y velocity
radius	float64	m	agent collision radius
planner	string	-	pedestrian motion model (Hive partition column)
scenario	string	-	canonical scenario id (Hive partition column)
seed	int64	-	RNG seed for the trial
robot_policy	string	-	robot policy name; empty string in divergence-mode trials
bucket	string	-	scenario family: nav / bt (Hive partition column)
density	string	-	pedestrian density regime: sparse / dense (Hive partition column)
modality	string	-	scenario modality within (bucket, density), e.g. corridor, group_conversation
is_robot_scenario	bool	bool	true if scenario id is a robot variant of a pure-ped sibling
source_dir	string	-	originating sweep dir name (provenance)

kinematics_per_trial

field	dtype	unit	description
bucket	string	-	scenario family: nav / bt (Hive partition column)
scenario	string	-	canonical scenario id (Hive partition column)
planner	string	-	pedestrian motion model (Hive partition column)
robot_policy	string	-	robot policy name; empty string in divergence-mode trials
seed	int64	-	RNG seed for the trial
jerk	float64	m/s^3	mean per-agent jerk over the trial
curvature	float64	1/m	mean per-agent path curvature over the trial
collisions	int64	count	agent-agent collision events in the trial

robot_metrics

field	dtype	unit	description
scenario	string	-	canonical scenario id (Hive partition column)
bucket	string	-	scenario family: nav / bt (Hive partition column)
ped_planner	string	-	pedestrian planner used; alias of planner in metrics tables
robot_policy	string	-	robot policy name; empty string in divergence-mode trials
seed	int64	-	RNG seed for the trial
source_dir	string	-	originating sweep dir name (provenance)
success	float64	0/1	robot reached its goal within tolerance
time_to_goal_s	float64	s	robot wall-time to first goal arrival; NaN if no arrival
ttg_ratio	float64	ratio	time-to-goal divided by straight-line nominal time
path_efficiency	float64	ratio	robot arc-length / start-to-goal straight-line distance
n_robot_collisions	float64	count	frames in which robot overlapped any human
personal_space_violations	float64	count	frames in which a human entered the robot's personal-space disk
psv_per_sec	float64	1/s	personal_space_violations normalized by trial duration
frozen_at_end	float64	0/1	robot mean speed in last freeze-window seconds is below threshold
final_goal_dist	float64	m	robot-goal distance at trial end

failures

field	dtype	unit	description
scenario	string	-	canonical scenario id (Hive partition column)
bucket	string	-	scenario family: nav / bt (Hive partition column)
ped_planner	string	-	pedestrian planner used; alias of planner in metrics tables
robot_policy	string	-	robot policy name; empty string in divergence-mode trials
seed	int64	-	RNG seed for the trial
source_dir	string	-	originating sweep dir name (provenance)
cause	string	-	failure-cause label (success / collision / timeout_no_progress / ...)

Quick start

from datasets import load_dataset

# trajectories - stream a single (bucket, density, scenario, planner) partition
ds = load_dataset("arena-rosnav/PEDS37", "agent_states", split="train", streaming=True,
                  data_files="data/agent_states/bucket=nav/density=sparse/scenario=nav_sparse_corridor/planner=sfm/*.parquet")

# or all of one bucket (partition pushdown)
ds_nav = load_dataset("arena-rosnav/PEDS37", "agent_states", split="train", streaming=True,
                      data_files="data/agent_states/bucket=nav/**/*.parquet")

# per-trial scalar metrics
import pandas as pd
kin = pd.read_parquet("hf://datasets/arena-rosnav/PEDS37/data/metrics/kinematics_per_trial.parquet")
print(kin.groupby("planner")[["jerk", "curvature", "collisions"]].mean())

Citation

Paper: PEDS: Closed-Loop Benchmarking of Social Navigation across Pedestrian Behavior Models, Duc Tai Nguyen, Volodymyr Shcherbyna, Anh Do Duc, Zhengcheng Shen, and Linh Kästner, IEEE Robotics and Automation Letters, 2026 (DOI to appear).

@article{nguyen2026peds,
  title   = {{PEDS}: Closed-Loop Benchmarking of Social Navigation across Pedestrian Behavior Models},
  author  = {Nguyen, Duc Tai and Shcherbyna, Volodymyr and Do Duc, Anh and Shen, Zhengcheng and K{\"a}stner, Linh},
  journal = {IEEE Robotics and Automation Letters},
  year    = {2026},
  note    = {DOI to appear}
}

License

CC BY 4.0 for the dataset.