thatsmee/wd2-keyboard-driving

Watch Dogs 2 Keyboard Driving

Human driving in Watch Dogs 2 (streamed via GeForce NOW on macOS), recorded as screen frames plus the exact keyboard state. Made for imitation learning / behaviour cloning with a discrete keyboard action space (W/A/S/D/Space).

• 187,295 labelled frames at 1280×720 from 27 recording sessions
• every session is one complete navigation run: driving from the start point along the in-game minimap route all the way to the destination, ending when the car is parked
• per-frame keyboard state and discrete action labels with nanosecond timestamps
• raw keyboard event logs per session (every key down/up), so you can build your own event-aware targets
• human annotations of road-branch affordances and minimap route geometry

The recordings come in two storage formats, kept exactly as recorded (no re-encoding):

Subset 1 – frames (JPEG, loadable with datasets)

WebDataset shards in data/. Each sample key is <session_id>_f<frame_id> with a .jpg (JPEG quality 0.9) and a .json. Splits are by whole session.

Split	Sessions	Frames	Driving time	Shards	Steer L / straight / R	Throttle	Brake/reverse
train	12	93,543	80 min	20	7.1 % / 86.6 % / 6.3 %	44.4 %	2.7 %
validation	2	12,365	10 min	3	5.8 % / 87.2 % / 7.0 %	43.8 %	4.6 %
test	3	21,805	20 min	5	5.7 % / 88.5 % / 5.8 %	37.2 %	2.1 %

The test split is three clean sessions recorded on 2026-07-17. The author evaluated one model on them once.

from datasets import load_dataset
ds = load_dataset("thatsmee/wd2-keyboard-driving", "frames", split="train", streaming=True)
sample = next(iter(ds))   # {'jpg': PIL.Image, 'json': {...}, '__key__': ...}

Subset 2 – videos (HEVC, fixed routes)

Sessions recorded on 2026-07-19 on fixed routes (route_id in metadata/sessions.jsonl), stored as 60-second HEVC .mov chunks in videos/<split>/<session_id>/. Per-frame labels are in metadata/video_frames/<session_id>.jsonl: file is the chunk path in this repo and chunk_frame_index the frame index inside that chunk. Most human annotations refer to these sessions.

Split	Sessions	Frames	Driving time	Steer L / straight / R	Throttle	Brake/reverse
train	6	29,740	24 min	5.7 % / 89.2 % / 5.2 %	43.9 %	2.1 %
validation	4	29,825	25 min	5.9 % / 89.4 % / 4.6 %	43.7 %	2.4 %

Decode e.g. with PyAV or ffmpeg -i chunk_00000.mov frames/%06d.jpg, then join on chunk_frame_index.

Labels (both subsets)

• action: steer (-1 left, 0 straight, 1 right), throttle, brake_reverse, handbrake (0/1)
• keys: raw pressed state of w a s d space at frame time
• flags: burst_capture, steer_conflict (A and D both down), duplicate_candidate, focus_valid, window_valid
• frame_id, frame_receipt_ts_ns, frame_pts_ns, visual_hash, width, height, session_id, split

Frames are not evenly spaced (burst saving around key events, near-duplicate skipping). Use timestamps, not indices, for timing.

Metadata

• metadata/sessions.jsonl – capture settings and statistics per session
• metadata/key_events/<session_id>.jsonl – every key down/up event (event_ts_ns, receipt_ts_ns, full key state)
• metadata/annotations/ – human labels, joined by session_id + frame_id:

File	Rows
minimap-route-human-geometry-500-clean.jsonl	500
minimap-route-human-geometry-review-500-arrival-relabels.jsonl	35
minimap-route-human-geometry-review-500.jsonl	500
minimap-route-human-geometry-review.jsonl	24
minimap-route-human-review.jsonl	24
onset-route-role-review-120.jsonl	6
paired-route-branch-150.jsonl	150
paired-route-branch-geometry-300.jsonl	300
paired-route-branch-random-150.jsonl	150
paired-route-branch-targeted-right-180.jsonl	180
screen-branch-affordance-500.jsonl	500
screen-branch-affordance-historical-500.jsonl	500
screen-branch-affordance-right-targeted-250.jsonl	250
screen-branch-affordance-temporal-150.jsonl	150
v8-temporal-braking-review.jsonl	16
v9-maneuver-human-review.jsonl	1
v9-onset-human-review.jsonl	34
v9-segment-human-review.jsonl	64
vision-first-route-hint-500.jsonl	500
vision-first-route-hint-right-targeted-300.jsonl	300

Known properties (read before training)

• Heavy class imbalance. Most frames are "straight"; right turns are clearly rarer than left turns; brake is rare and noisy.
• Copy-the-last-action shortcut. A model that sees past actions learns to repeat the previous key. It scores well on frame-level F1 (repeating the last expert action matches it) but collapses in closed-loop driving. Evaluate onset/transition metrics.
• Vision-only baseline from the author: steering macro-F1 ≈ 0.38 vs ≈ 0.31 for always-straight; brake F1 near 0.

License and content

Recordings, labels and annotations: CC BY-NC 4.0 (non-commercial). The visual content of the frames is from Watch Dogs 2, © Ubisoft Entertainment. This is an unofficial, non-commercial research dataset, not affiliated with or endorsed by Ubisoft or NVIDIA. Rights holders who want content removed can open a discussion on this repository.