StarLineResearch/vehicle-control-dataset

Vehicle Control Dataset

Overview

• Data type: Vehicle telemetry (temporal actions + tracking errors)
• Size: 58,862 rows, 15 columns
• Coverage: 18 driving days over 191 calendar days (Apr — Nov 2024)
• Format: Parquet + CSV

Schema

Column	Type	Description
series_id	int64	Session identifier (bag_id → int64)
timestamp	int64	Nanoseconds since Unix epoch
velocity	float	Vehicle speed (m/s)
steering	float	Steering wheel angle (rad)
gas	float	Gas pedal position [0, 1]
brake	float	Brake pedal position [0, 1]
acceleration	float	Longitudinal acceleration (m/s²)
acceleration_x	float	Acceleration X component (m/s²)
acceleration_y	float	Acceleration Y component (m/s²)
lateral_error	float	Cross-track error (m)
longitudinal_error	float	Along-track error (m)
heading_error	float	Heading error (rad)
error_energy	float	Integrated error energy
goal_x	float	Relative goal position X (m)
goal_y	float	Relative goal position Y (m)

Goal Computation

The goal vector is computed by rotating the tracking error vector from the Frenet frame of the reference trajectory into global coordinates:

θ = heading_v - heading_error

goal_x = longitudinal_error · cos(θ) - lateral_error · sin(θ)
goal_y = longitudinal_error · sin(θ) + lateral_error · cos(θ)

where heading_v is the vehicle heading derived from consecutive GPS measurements (UTM → atan2).

Goals are only available when both errors and GPS heading are non-null.

Coverage

Column	Non-null	%
series_id, timestamp	58,862	100.0%
longitudinal_error, heading_error, error_energy	57,691	98.0%
acceleration (3 axes)	56,839	96.6%
lateral_error	47,723	81.1%
velocity	38,895	66.1%
steering	27,075	46.0%
goal_x, goal_y	29,735	50.5%
gas, brake	~15,894	27.0%

Numeric Distribution

Column	min	median	mean	max	std
velocity	0.0	22.65	21.78	27.53	3.43
steering	−0.67	0.0	−0.001	0.70	0.03
gas	0.0	0.39	0.39	0.77	0.12
brake	0.0	0.0	0.002	0.80	0.02
acceleration	−5.81	0.01	0.002	1.99	0.16
lateral_error	−1.00	0.05	0.05	1.24	0.12
longitudinal_error	−5.5e5	−0.0	−3.9e3	1.1e5	3.3e4
heading_error	−3.09	−0.002	0.004	3.12	0.19
goal_x	−3.5e5	0.01	−21.1	0.53	2.6e3
goal_y	−0.99	0.01	21.4	3.5e5	2.6e3

Usage

import pandas as pd

df = pd.read_parquet("telemetry.parquet")
print(f"{len(df)} rows, {df['series_id'].nunique()} sessions")

# Train/val/test split by session
series_ids = df['series_id'].unique()
n = len(series_ids)
train_ids = series_ids[:int(n * 0.8)]
val_ids   = series_ids[int(n * 0.8):int(n * 0.9)]
test_ids  = series_ids[int(n * 0.9):]

train = df[df['series_id'].isin(train_ids)]
val   = df[df['series_id'].isin(val_ids)]
test  = df[df['series_id'].isin(test_ids)]

License

CC0 1.0 Universal — No rights reserved.

---

Source: Fleet of autonomous vehicles, real driving sessions

Sensor: GPS (NovAtel), IMU, CAN bus, control evaluator module

Coverage: 58,862 timesteps, 33 sessions, 191 calendar days (18 driving days)

Format: Parquet + CSV

Total file size: 3.2 MB (Parquet) + 11.9 MB (CSV)