ryangowe/ets2-dataset

ETS2 Truck Self-Driving Dataset

Synchronized screen-capture video and vehicle telemetry recorded from human driving in Euro Truck Simulator 2 (ETS2), packaged for end-to-end driving (imitation learning / behavioral cloning). Each frame pairs a 1080p game screenshot with the full SCS telemetry snapshot taken at the same instant — the control signals (steering, throttle, brake) and vehicle state needed to learn or evaluate a driving policy.

If you want to load a single recording locally, jump to Usage. If you want to understand what each field means, see Telemetry fields.

At a glance

• Modalities: RGB video + per-frame structured telemetry
• Frame rate: 10 FPS, video and telemetry aligned 1:1
• Frame size: 1920×1080 (H.264, yuv420p)
• Format: WebDataset — .tar shards of ~1 GB each
• Splits: train, val

Dataset structure

The dataset is a set of WebDataset shards named ets2-{split}-{index}.tar (e.g. ets2-train-000000.tar). Each shard is a plain tar of complete recording sessions. A session is one continuous driving clip (15–90 s) stored as three members that share a basename (the WebDataset key); the extension names the field:

Member	Contents
{key}.mp4	H.264 video, 1920×1080, 10 FPS
{key}.bin.zst	zstd-compressed raw telemetry: a fixed 32 KiB SCS shared-memory block per frame, concatenated in frame order
{key}.json	session meta needed to interpret the telemetry (see Session meta)

The telemetry bin is the only compressed member. After zstd decompression it is 32768 × N bytes for an N-frame video; frame i corresponds to bytes [i*32768 : (i+1)*32768], the verbatim SCS plugin shared memory at that frame.

Telemetry fields

Fields are decoded from the 32 KiB block by byte offset. The block is stored verbatim, so it carries the entire SCS shared-memory layout; the fields below are the ones already wired up for driving. Angles follow the SCS convention of turns (1 turn = 360°).

Field	Type	Meaning
game_steer	float	Effective steering, [-1, 1]
game_throttle	float	Effective throttle, [0, 1]
game_brake	float	Effective brake, [0, 1]
speed	float	Truck speed in m/s (negative when reversing)
cruise_control_speed	float	Cruise-control set speed in m/s (0 when off)
angular_velocity_y	float	Yaw rate around truck-Y in turns/s
cabin_aa_z	float	Cabin angular acceleration around truck-Z in turns/s²
coordinate_x / coordinate_y / coordinate_z	float	World position (ETS2 left-handed frame)
rotation_x / rotation_y / rotation_z	float	Heading / pitch / roll in turns (heading 0 = north)
time_ms	int	Simulation time in ms; stops while the game is paused
simulated_time_ms	int	Simulation time in ms; keeps advancing while paused
sdk_active / paused / on_job / attached	bool	SDK active, game paused, delivery job active, trailer attached
wear_engine / wear_transmission / wear_cabin / wear_chassis / wear_wheels	float	Component wear, [0, 1]
cargo_damage	float	Current cargo damage, [0, 1]
route_distance	float	Remaining planned route distance in m
planned_distance_km	int	Initial route length in km (constant for the trip)
job_income	int	Job payment in EUR (constant for the trip)
plugin_revid / sdk_version_major / sdk_version_minor	int	Telemetry plugin / SDK versions

For driving you typically use the screenshot plus speed and pose as inputs and game_steer / game_throttle / game_brake as labels.

Session meta

{key}.json is a small object describing how to read the telemetry bin:

Key	Meaning
plugin_revid	Telemetry plugin revision (struct layout version)
sdk_version_major / sdk_version_minor	SCS SDK version
hfov_deg	Camera horizontal field of view in degrees (default 71)

Data collection

Sessions are captured from a live ETS2 window, not reconstructed offline:

• Sampling: the game window is screen-captured at 10 Hz. Accepted window sizes are 1920×1080, 2560×1440, and 3840×2160; every frame is normalized to 1920×1080. Each screenshot is paired with the SCS telemetry snapshot read at the same tick.
• Event-driven recording. Frames are buffered with a few seconds of look-ahead, and a clip is only kept around interesting moments rather than recording idle time. Two triggers are active:
  • moving — fires while the truck moves faster than 0.3 m/s, keeping a [-3 s, +1 s] window around each moving frame;
  • collision — fires when total component wear jumps between frames (a likely impact), keeping a wider [-3 s, +12 s] window.
• Clip length. Kept sessions are 15–90 s; shorter candidate clips are discarded and longer runs are split.

Usage

Stream shards directly

WebDataset shards stream without downloading the whole dataset. The telemetry member needs zstd decompression, then slicing into 32 KiB per-frame blocks:

import webdataset as wds
import zstandard as zstd

REPO = "<your-username>/<dataset-repo>"  # replace with the dataset repo id
BASE = f"https://huggingface.co/datasets/{REPO}/resolve/main"
url = f"{BASE}/ets2-train-{{000000..000006}}.tar"  # adjust to the shard range

FRAME_BYTES = 32 * 1024

for sample in wds.WebDataset(url):
    meta = sample["json"]              # bytes — session meta
    video = sample["mp4"]              # bytes — H.264 clip
    raw = zstd.ZstdDecompressor().decompress(sample["bin.zst"])
    n_frames = len(raw) // FRAME_BYTES
    # frame i telemetry: raw[i*FRAME_BYTES : (i+1)*FRAME_BYTES]

Decode with the recorder toolkit

This dataset is produced by the ets2-dataset tooling in this repository, which can also read sessions back. Unpack a shard (plain tar -xf) and open any session by pointing at one of its members:

from ets2_dataset.data.session import Session

with Session("session_20260101_120000_123456.mp4") as session:
    print(len(session), "frames", session.fps, "fps", session.duration, "s")
    for frame in session:
        image = frame.image                 # [H, W, 3] BGR uint8
        t = frame.telemetry
        steer, throttle, brake = t.game_steer, t.game_throttle, t.game_brake

Session decodes the video and telemetry together and yields frames in recording order, with telemetry exposing the fields above as attributes.

License and attribution

The video frames are screenshots of Euro Truck Simulator 2, whose game content is © SCS Software. This dataset is intended for non-commercial research and is distributed under that constraint; using it does not grant any rights to the underlying game assets. Set the license field in the metadata above to the terms you intend to release the recorded telemetry and annotations under before publishing.