Vehicle_Power_Type_Trajectory_Dataset (VTT)

Can you tell what kind of vehicle produced a trace from how it moves?

VTT is a sequence-classification dataset of trajectory and CAN-bus signals at 10 Hz: ego speed and acceleration, steering angle, steering rate and normalized steering torque, body yaw rate, pedal and turn-signal flags, and the interaction with the lead vehicle (distance, relative speed, lead acceleration, headway, time to collision). Each drive is labelled with the footprint size class and the powertrain type of the vehicle that produced it. The release contains 9,062 drives (3,949.8 h) from 1,431 pseudonymous vehicles on 360 vehicle platforms of 38 brands, and 201,390 eligible 60 s windows. Periods during which a driving automation system (DAS) controlled the vehicle are kept and labelled per sample (62.6 % of all samples show automation activity), so the data describe both human driving and automated driving of the same vehicles.

No video, images, positions, timestamps or driver-monitoring signals are included.

At a glance

Drives (one ignition cycle each)	9,062
Pseudonymous vehicles (device x platform)	1,431
Identity groups (unit of all splits)	1,157
Vehicle platforms / brands / models	360 / 38 / 174
Hours (nominal) / hours moving	3,949.8 / 3,078.1
Share of samples with DAS engaged / with any automation activity	47.9 % / 62.6 %
60 s windows: listed / eligible / human-only (strict)	241,476 / 201,390 / 37,942
Eligible windows per task: size / BEV-vs-ICE / four-class powertrain	201,169 / 144,534 / 178,827
Windows in the ready-to-train shards	194,949 in 41 files
Size on disk: drive tables / window shards	8.1 GB / 4.4 GB

Labels

Footprint size class (task A)

size_class is derived from the vehicle platform's footprint = wheelbase x mean track width (m^2) with two fixed thresholds: SMALL below 4.43 m^2, MID up to 4.88 m^2, LARGE above. The underlying specifications (wheelbase_mm, track_front_mm, track_rear_mm, footprint_m2, curb_mass_kg, body, the legacy segment code size_code) are published per drive so that other size definitions can be derived. size_class is null when the platform's specifications are unknown; such drives are not eligible for task A (task_size_eligible = False).

Powertrain (task B)

powertrain is one of ICE, HEV, PHEV, BEV, or null. powertrain_raw keeps the seven-value platform code (ICE, HEV, PHEV, BEV, MIXED, UNK, FCEV); platforms sold with several powertrains that cannot be told apart from the logs are MIXED and get a null powertrain. The main task B is BEV vs ICE (task_bev_ice_eligible = powertrain in {BEV, ICE}); the four-class problem (powertrain not null) is a secondary task. The two tasks are evaluated separately.

Class support (eligible windows)

task	class	eligible windows	hours	vehicles	platforms
size	SMALL	91,369	1,522.8	673	169
size	MID	56,059	934.3	392	104
size	LARGE	53,741	895.7	295	76
size	null	221	3.7	10	6
powertrain	ICE	79,566	1,326.1	598	200
powertrain	HEV	23,708	395.1	197	57
powertrain	PHEV	10,585	176.4	68	21
powertrain	BEV	64,968	1,082.8	366	71
powertrain	null	22,563	376.1	156	34

Splits

Every drive carries two independent split assignments. The unit of both is the identity group (group_id: same device, same vehicle identity or shared drive names), so no group appears in two splits.

• split_platform (headline): whole car models (the lab taxonomy's base models, column model) are held out as the test split; a held-out model never appears in training or validation (about 15 % of the hours of each size x powertrain cell, 41 held-out models, chosen so that a brand contributes both powertrain classes where possible). The remaining groups are split group-disjoint into train and validation (about 15 % of the remaining hours), so a model may be shared between training and validation. Drives of a group that also drove a held-out model are marked excluded on its other models: they belong to no split and are left out of every protocol.
• split_device (secondary): groups are assigned at random within each size x powertrain cell with 70 / 15 / 15 % of the eligible hours; models may be shared across splits.

Lab-owned devices are always in train. A size x powertrain cell with fewer than 10 test hours, 5 groups or 2 platforms is flagged low_support_cell and is reported separately, never inside headline numbers.

task	split column	split	eligible windows	hours	groups	platforms
size	split_platform	train	143,400	2,390.0	672	216
size	split_platform	validation	25,188	419.8	137	85
size	split_platform	test	26,144	435.7	297	102
size	split_device	train	142,614	2,376.9	724	298
size	split_device	validation	29,341	489.0	193	125
size	split_device	test	29,214	486.9	191	120
bev_ice	split_platform	train	102,817	1,713.6	479	168
bev_ice	split_platform	validation	18,564	309.4	89	55
bev_ice	split_platform	test	18,295	304.9	222	77
bev_ice	split_device	train	101,633	1,693.9	534	230
bev_ice	split_device	validation	21,451	357.5	131	89
bev_ice	split_device	test	21,450	357.5	131	84

Protocols

All filters are boolean expressions over index/windows_60s.parquet columns (pandas syntax). Replace task_size_eligible / size_class by task_bev_ice_eligible / powertrain for task B.

protocol	filter
train	eligible & task_size_eligible & (split_platform == "train")
validation	eligible & task_size_eligible & (split_platform == "validation")
test_full (headline)	eligible & task_size_eligible & (split_platform == "test")
test_balanced	window_id.isin(splits/eval_balanced_size_test.parquet.window_id) (class-balanced, at most 25 % of a class from one group and 50 % from one car model)
test_human_only	test_full & human_only_strict (no automation activity in the window nor in the 10 s before it)
drive_full	one prediction per test drive: the mean of the class-probability vectors of the drive's test_full windows, arg-max (a tie goes to the earlier class of the task); drives without an eligible window are excluded and counted separately
drive_human_only	the same over the drive's test_human_only windows
vehicle_full	one prediction per pseudonymous vehicle: the mean over the test_full windows of all its test drives; the subset of vehicles with at least two drives is reported as vehicle_full_multi
split_device variants	the same filters with split_device in place of split_platform; the balanced lists of this scheme are splits/eval_balanced_<task>_test_device.parquet

Headline metric: macro-averaged F1 over the classes of the task on test_full, excluding windows with low_support_cell = True (reported separately). splits/train_uniform_<task>.parquet provides a per-window weight (inverse class frequency to the power 0.5, capped at 10) for class-balanced training. The balanced evaluation lists and both splits were generated with seed 20261105.

Official model inputs

Under the official protocol a model may use only the following drive-table columns (the same columns form the 600-step sequences of the window shards). Everything else in the drive tables and all index columns are metadata for analysis and must not be fed to a model (they carry brand- and platform-specific shortcuts).

official input	unit	meaning
v_mps	m/s	ego speed
a_mps2	m/s^2	ego longitudinal acceleration (vehicle estimate)
steer_angle_deg	deg	steering wheel angle, left positive
steer_rate_degps	deg/s	steering wheel angle rate
steer_torque_norm	-	steer_torque_raw divided by the drive's 95th percentile of |torque| while moving and steering
yaw_rate_radps	rad/s	body yaw rate, left positive (device IMU / localizer)
curv_1pm	1/m	path curvature = yaw_rate / v (v >= 1 m/s and yaw_valid, else NaN)
lat_acc_mps2	m/s^2	lateral acceleration = v * yaw_rate
gas_pressed	-	accelerator pedal pressed
brake_pressed	-	brake pedal pressed
steer_pressed	-	driver torque above the vehicle's threshold
standstill	-	vehicle reports standstill
das_enabled	-	driving automation engaged: controller enabled OR stock cruise enabled
lead_valid	-	lead vehicle tracked (status and distance > 0.5 m, sample fresh <= 0.75 s)
lead_d_m	m	longitudinal distance to lead
lead_v_rel_mps	m/s	lead speed minus ego speed
lead_speed_mps	m/s	lead speed = v_mps + lead_v_rel_mps
lead_a_slope_mps2	m/s^2	lead acceleration: causal 1 s least-squares slope of lead speed within the lead episode
thw_s	s	time headway = lead_d_m / v_mps (v > 0.5 m/s)
ttc_s	s	time to collision = lead_d_m / closing speed (closing > 0.1 m/s)

Not model inputs: seg, gap, v_raw_mps, steer_torque_raw, yaw_valid, pose_source, gas, brake, blinker_left, blinker_right, das_any, op_lat_active, op_long_active, cruise_available, v_set_mps, lead_y_m, lead_source, lead_prob, lead2_valid, lead2_d_m, lead2_v_rel_mps, qc_lead_age_s, qc_yaw_age_s, and every column of the index tables.

Reference results

Classifiers that use the official channels only (hand-crafted window statistics with logistic regression and gradient boosting; a one-dimensional ResNet, a bidirectional GRU, a Transformer encoder and Mamba state-space models) are reported with their code in experiments/RESULTS.md of the code repository, under exactly these protocols and split schemes. They are reference points for the two tasks, not a leaderboard.

Files

README.md                      this card
LICENSE                        CC BY 4.0
CHECKSUMS.sha256               sha256 of every other file (sha256sum -c compatible)
index/drives.parquet           one row per drive: labels, specifications, split assignments, provenance flags
index/vehicles.parquet         one row per pseudonymous vehicle
index/platforms.parquet        one row per vehicle platform with its specifications
index/windows_60s.parquet      one row per window (eligibility, statistics, labels, splits)
index/split_summary.csv        hours / windows / groups / platforms per task, class and split
index/readme_numbers.json      every number printed on this card
splits/splits.json             split configuration, per-cell support, group assignments, held-out platforms
splits/eval_balanced_<task>_<split>.parquet   window ids of the balanced evaluation lists (split_platform scheme)
splits/eval_balanced_<task>_<split>_device.parquet   the same lists for the split_device scheme
splits/train_uniform_<task>.parquet           per-window training weights (split_platform scheme)
splits/train_uniform_<task>_device.parquet    the same weights for the split_device scheme
schema/*.json                  machine-readable column descriptions (name, type, unit, description, source)
data/vehicle_NNNN/drive_NNNNNN.parquet        full drive tables, 10 Hz, one row per sample
windows/{train,validation,test}/*.parquet     ready-to-train window shards (index columns + sequences)

Drive tables (data/)

Row k is the k-th sample of the vehicle-state stream; t_s = k / 10 is nominal time. A gap flag marks rows preceded by more than 0.3 s without data. Lower-rate streams (lead vehicle, yaw rate, automation state) are held causally to the sample times and set to NaN when older than 0.75 s (lead) or 0.625 s (yaw).

column	type	unit	description	source field	role
t_s	float64	s	nominal time from the first sample of the drive, = row_index / 10	row index	metadata
seg	int16	-	ordinal of the source 1-min segment within the drive (0-based)	segment number	metadata
gap	bool	-	True when more than 0.3 s elapsed since the previous sample (missing data before this row)	sample times	metadata
v_mps	float32	m/s	ego speed	carState.vEgo	input
v_raw_mps	float32	m/s	unfiltered ego speed	carState.vEgoRaw	metadata
a_mps2	float32	m/s^2	ego longitudinal acceleration (vehicle estimate)	carState.aEgo	input
standstill	bool	-	vehicle reports standstill	carState.standstill	input
steer_angle_deg	float32	deg	steering wheel angle, left positive	carState.steeringAngleDeg	input
steer_rate_degps	float32	deg/s	steering wheel angle rate	carState.steeringRateDeg	input
steer_torque_raw	float32	raw	driver steering torque in the vehicle's own CAN units (not comparable across makes)	carState.steeringTorque	metadata
steer_torque_norm	float32	-	steer_torque_raw divided by the drive's 95th percentile of |torque| while moving and steering	derived	input
steer_pressed	bool	-	driver torque above the vehicle's threshold	carState.steeringPressed	input
yaw_rate_radps	float32	rad/s	body yaw rate, left positive (device IMU / localizer)	livePose or liveLocationKalman angularVelocityDevice.z	input
yaw_valid	bool	-	yaw sample fresh (<= 0.625 s) and flagged valid by the localizer	derived	metadata
pose_source	int8	-	0 none, 1 livePose (5 Hz), 2 liveLocationKalman (4 Hz)	derived	metadata
curv_1pm	float32	1/m	path curvature = yaw_rate / v (v >= 1 m/s and yaw_valid, else NaN)	derived	input
lat_acc_mps2	float32	m/s^2	lateral acceleration = v * yaw_rate	derived	input
gas	float32	0-1	accelerator pedal position (NaN on drives where the channel is dead)	carState.gas	metadata
gas_pressed	bool	-	accelerator pedal pressed	carState.gasPressed	input
brake	float32	0-1	brake pedal position / pressure (NaN on drives where the channel is dead)	carState.brake	metadata
brake_pressed	bool	-	brake pedal pressed	carState.brakePressed	input
blinker_left	bool	-	left turn signal	carState.leftBlinker	metadata
blinker_right	bool	-	right turn signal	carState.rightBlinker	metadata
das_enabled	bool	-	driving automation engaged: controller enabled OR stock cruise enabled	controlsState.enabled | carState.cruiseState.enabled	input
das_any	bool	-	das_enabled OR any lateral/longitudinal automation activity flag	das_enabled | selfdriveState.enabled | carControl.enabled | latActive | longActive	metadata
op_lat_active	bool	-	automation steering active	carControl.latActive	metadata
op_long_active	bool	-	automation longitudinal control active	carControl.longActive	metadata
cruise_available	bool	-	cruise control available	carState.cruiseState.available	metadata
v_set_mps	float32	m/s	cruise set speed (NaN when unset)	carState.cruiseState.speed / vCruise	metadata
lead_valid	bool	-	lead vehicle tracked (status and distance > 0.5 m, sample fresh <= 0.75 s)	radarState.leadOne	input
lead_d_m	float32	m	longitudinal distance to lead	radarState.leadOne.dRel	input
lead_y_m	float32	m	lateral offset of lead	radarState.leadOne.yRel	metadata
lead_v_rel_mps	float32	m/s	lead speed minus ego speed	radarState.leadOne.vRel	input
lead_speed_mps	float32	m/s	lead speed = v_mps + lead_v_rel_mps	derived	input
lead_a_slope_mps2	float32	m/s^2	lead acceleration: causal 1 s least-squares slope of lead speed within the lead episode	derived	input
lead_source	int8	-	0 none, 1 radar track, 2 camera model track	radarState.leadOne.radar / carParams.radarUnavailable	metadata
lead_prob	float32	-	model probability of the lead (NaN on older software)	radarState.leadOne.modelProb	metadata
thw_s	float32	s	time headway = lead_d_m / v_mps (v > 0.5 m/s)	derived	input
ttc_s	float32	s	time to collision = lead_d_m / closing speed (closing > 0.1 m/s)	derived	input
lead2_valid	bool	-	second lead tracked	radarState.leadTwo	metadata
lead2_d_m	float32	m	distance to second lead	radarState.leadTwo.dRel	metadata
lead2_v_rel_mps	float32	m/s	second lead relative speed	radarState.leadTwo.vRel	metadata
qc_lead_age_s	float32	s	age of the held lead sample (provenance, not a model input)	derived	metadata
qc_yaw_age_s	float32	s	age of the held yaw sample (provenance, not a model input)	derived	metadata

Window shards (windows/)

One row per eligible window that is eligible for at least one task, assigned to train / validation / test by split_platform. Columns: every column of index/windows_60s.parquet plus one 600-element float32 sequence per official input (booleans as 0 / 1). Files are at most 500 MB each.

split	files	windows	GB
test	6	26,163	0.6
train	29	143,591	3.2
validation	6	25,195	0.6

Index tables

index/drives.parquet

column	type	description
vehicle_id	str	pseudonymous vehicle (device x platform), random order
drive_id	str	pseudonymous drive (one ignition cycle / openpilot route), random order
group_id	str	identity cluster used as the unit of all splits
file	str	relative path of the drive table
n_rows	int32	rows of the drive table (10 Hz samples)
duration_s	float32	nominal duration = n_rows / 10
n_segments	int16	source segments
n_seg_rlog_dec	int16	segments decimated from the high-rate log (no low-rate log present)
n_gaps	int16	rows flagged gap
max_gap_s	float32	largest gap
brand	str	manufacturer
model	str	base model
platform	str	canonical vehicle platform (generation) name
body	str	body style
size_code	str	legacy US segment code
wheelbase_mm	float32	wheelbase
track_front_mm	float32	front track width
track_rear_mm	float32	rear track width
footprint_m2	float32	wheelbase x mean track
curb_mass_kg	float32	curb mass (representative)
size_class	str	SMALL / MID / LARGE by footprint thresholds (null when unknown)
powertrain	str	ICE / HEV / PHEV / BEV (null when the platform is shared by several powertrains)
powertrain_raw	str	7-value platform code incl. MIXED / UNK / FCEV
label_source	str	how the platform was identified
label_note	str	known label-noise notes
region	str	north_america / europe / asia / other_unknown
software_era	str	coarse software generation
device_type	str	logging device hardware class
pose_source	str	livePose / liveLocationKalman / mixed / none
radar_unavailable	bool	lead comes from the camera model (True) or radar (False)
op_long	bool	automation stack controlled longitudinal motion when engaged
gas_alive	bool	gas magnitude channel populated
brake_alive	bool	brake magnitude channel populated
torque_alive	bool	steering torque channel populated
has_pose	bool	yaw-rate stream present
moving_s	float32	seconds with v >= 1 m/s
das_s	float32	seconds with das_enabled
das_any_s	float32	seconds with das_any
human_s	float32	seconds without das_any
lead_valid_frac	float32	share of samples with a valid lead
cf_frac	float32	share of valid-lead samples with THW <= 6 s
yaw_valid_frac	float32	share of samples with a valid yaw sample
steer_torque_scale	float32	divisor used for steer_torque_norm
split_device	str	train / validation / test (device-disjoint)
split_platform	str	train / validation / test / excluded (headline: test platforms never appear in train or validation; excluded = the drive's group also drove a held-out platform, the drive belongs to no split)
task_size_eligible	bool	size_class known
task_bev_ice_eligible	bool	powertrain in {BEV, ICE}
low_support_cell	bool	cell has too few groups / platforms / hours for headline metrics

index/windows_60s.parquet

A window is eligible when it has 600 consecutive samples, no gap, and the vehicle is moving (v >= 1 m/s) for at least 20 % of the samples.

column	type	description
window_id	str	_w
drive_id	str	-
vehicle_id	str	-
group_id	str	-
w_idx	int32	window index within the drive
row0	int32	first row (inclusive)
row1	int32	last row (exclusive)
t0_s	float32	nominal start time
span_s	float32	nominal span of the window's samples = (rows - 1) x 0.1 s (59.9 s for a full window); the 60-s guarantee is gap_count == 0
eligible	bool	600 samples, no gap row inside the window, moving fraction >= 0.2
inelig_reason	str	-
kind	str	human / das / mixed by das_any
das_frac	float32	-
das_any_frac	float32	-
op_long_frac	float32	-
op_lat_frac	float32	-
moving_frac	float32	-
standstill_frac	float32	-
gap_count	int16	-
v_mean	float32	-
v_max	float32	-
v_std	float32	-
a_min	float32	-
a_max	float32	-
a_std	float32	-
n_accel_ep	int16	acceleration episodes: monotone runs of the smoothed speed (0.3 m/s tolerance) lasting >= 3 s with dv >= 2 m/s and mean acceleration >= 0.1 m/s^2; a run includes the steady cruise around a manoeuvre, which dilutes the mean of a short acceleration inside a long cruise
n_decel_ep	int16	deceleration episodes, same rule with the sign reversed
lead_frac	float32	-
cf_frac	float32	THW <= 6 s share of valid-lead samples
thw_med	float32	-
lead_speed_mean	float32	-
n_lead_accel_events	int16	-
yaw_valid_frac	float32	-
yaw_abs_mean	float32	-
curv_abs_mean	float32	-
steer_pressed_frac	float32	-
gas_pressed_frac	float32	-
brake_pressed_frac	float32	-
human_only	bool	no das_any sample in the window
human_only_strict	bool	human_only and no das_any in the preceding 10 s
size_class	str	-
powertrain	str	-
powertrain_raw	str	-
platform	str	-
brand	str	-
model	str	-
body	str	-
size_code	str	-
footprint_m2	float32	-
split_device	str	-
split_platform	str	-
task_size_eligible	bool	-
task_bev_ice_eligible	bool	-
low_support_cell	bool	-

index/vehicles.parquet

column	type	description
vehicle_id	str	pseudonymous vehicle (device x platform)
group_id	str	identity cluster (unit of all splits)
platform	str	canonical vehicle platform
brand	str	-
model	str	-
body	str	-
size_class	str	-
powertrain	str	-
powertrain_raw	str	-
region	str	-
n_drives	int32	drives of this vehicle
duration_h	float32	nominal hours
moving_h	float32	hours with v >= 1 m/s
das_any_h	float32	hours with any automation activity
n_windows_eligible	int32	eligible windows
split_device	str	-
split_platform	str	-
low_support_cell	bool	-

index/platforms.parquet

column	type	description
platform	str	canonical vehicle platform
brand	str	-
model	str	-
body	str	-
size_code	str	-
wheelbase_mm	float32	-
track_front_mm	float32	-
track_rear_mm	float32	-
footprint_m2	float32	-
curb_mass_kg	float32	-
size_class	str	-
powertrain	str	-
powertrain_raw	str	-
n_vehicles	int32	-
n_groups	int32	-
n_drives	int32	-
duration_h	float32	-
n_windows_eligible	int32	-
split_platform	str	split of the platform's drives: train / validation / test; 'mixed' when its groups fall in several splits (validation is group-disjoint inside the non-held-out pool, so train+validation is expected; 'test' is exclusive), 'excluded' when every drive of the platform is excluded

Limitations and caveats

• Labels are platform attributes. size_class and powertrain describe the vehicle platform identified from the vehicle's own bus identification, not a measurement of each drive. A misidentified platform mislabels every drive of that vehicle; known label-noise notes are in label_note.
• MIXED platforms are null. Platforms sold with several powertrains that the logs cannot separate carry no powertrain label (11.1 % of the drives) and are excluded from task B.
• Mild hybrids follow the platform code of their generation and may be labelled ICE.
• Automation share. 62.6 % of the samples show DAS activity; control style during these periods reflects the controller as much as the vehicle. Use das_enabled, das_any, human_only and human_only_strict to restrict analyses.
• Lead-vehicle source. On 56.2 % of the drives the lead vehicle comes from a camera model rather than a radar (radar_unavailable, lead_source); distance noise and availability differ between the two.
• Torque units. steer_torque_raw is in each make's own CAN units and is not comparable across brands; steer_torque_norm divides by the drive's 95th percentile of |torque| while moving and steering.
• Dead channels. The pedal magnitudes are not populated on every vehicle: gas is NaN on 63.9 % and brake on 49.4 % of the drives (gas_alive, brake_alive); the *_pressed flags are always valid.
• Quantization and sampling. Signal resolution and the exact sample clock differ by make; tables are indexed by sample, not by a fixed clock, and t_s is nominal.
• Region and software confounds. region, software_era, device_type, pose_source and the lead source are correlated with brand and powertrain; they are metadata, not inputs.
• Low-support cells. Some size x powertrain cells have few groups or platforms; they are flagged low_support_cell and excluded from headline metrics.
• Duplicates. Drives stored twice were merged before pseudonymisation; drives of one device remain correlated, which is why groups, not drives, are the unit of the splits.
• Car-following definitions. thw_s requires v > 0.5 m/s; the car-following share cf_frac uses headway <= 6 s.

Privacy and ethics

• Vehicles, drives and groups carry random, independent pseudonyms drawn for this release only; they cannot be linked to identifiers of any other dataset.
• The release contains no positions, no absolute times or dates, no video or images, and no driver-monitoring signals. region is a continent-level bucket; seg is an ordinal; t_s is relative to the start of the drive.
• Labels describe vehicle models, not people. Rare models may be recognisable as a model, never as a person.
• Access is granted for research on vehicle dynamics and driving behaviour. Re-identification of vehicles or drivers and linkage with other datasets are prohibited.
• Questions, corrections or removal requests: henryyuhangwang@gmail.com.

Citation

If you use VTT, please cite the paper below (arXiv:2603.06986):

@article{wang2026adasto,
  title   = {ADAS-TO: A Large-Scale Multimodal Naturalistic Dataset and
             Empirical Characterization of Human Takeovers during ADAS Engagement},
  author  = {Wang, Yuhang and Xu, Yiyao and Sun, Jingran and Zhou, Hao},
  journal = {arXiv preprint arXiv:2603.06986},
  year    = {2026},
  url     = {https://arxiv.org/abs/2603.06986}
}

License

The data tables are released under CC BY 4.0 (see LICENSE). The build and evaluation code is released under the MIT license at WangYuHang-cmd/-Vehicle_Power_Type_Trajectory_Dataset. Dataset repository: HenryYHW/Vehicle_Power_Type_Trajectory_Dataset; release v1, table schema 1.0, build configuration vtt-1.0.