By routine

The same pick-one-see-what-it-does reference By set piece gives set pieces, for the layer that describes how a restart was taken – a rule-based taxonomy, a data-driven alternative, and two type-specific layers (throw-ins, penalties) that don’t fit the general shape.

Routine taxonomy

What it describes: A rule-based routine_type per delivery (short / central-six-yard / penalty-area / recycled / deep for corners, an equivalent taxonomy per type), plus geometry (distance, angle, progression, side/third/channel) and a stable routine_key. For corners/free kicks specifically: delivery_technique ("inswinger"/"outswinger", from Opta qualifiers 223/224) and post_target ("near_post"/"far_post"/"central").
Requirements: Any set-piece type – works generically, with extra columns unlocked for corner/free kick.

Compute it

from wa_setpieces import restart_routines, analyze_routines

restart_routines(events, "corner")
analyze_routines(events, "corner", min_taker_attempts=3)
# .detail / .summary / .team_profiles / .taker_profiles / .target_matrix

Where it’s used: run_workflow()’s routines/ routine_summary/routine_team_profiles/routine_taker_profiles/ routine_target_matrix fields, for every type in By set piece.

Data-driven clusters

What it describes: Deliveries grouped by geometric similarity (k-means over start/end x/y by default) instead of a fixed, hand-picked taxonomy – surfaces whatever patterns a team actually repeats.
Requirements: The ml extra (pip install "wa-setpieces[ml]", scikit-learn).

Compute it

from wa_setpieces import cluster_routines, cluster_summary

clustered = cluster_routines(events, "corner", n_clusters=5)
# restart_routines' detail plus `cluster` (int, -1 = unclustered) and
# an auto-generated `cluster_label` (e.g. "short, forward, central")
cluster_summary(clustered)  # per-cluster usage/outcome roll-up

Where it’s used: plot_routine_clusters() in the Gallery; run_workflow’s routine_clusters field.

Long throws

What it describes: The one throw-in pattern that plays like a corner – a direct ball into a crowded box, where a genuine long-throw specialist is a real tactical weapon. Per-player usage/threat, and flick-on/knockdown detection restricted to throws past a distance threshold.
Requirements: throw_in only.

Compute it

from wa_setpieces import long_throw_taker_summary, long_throw_second_phases

long_throw_taker_summary(events, min_distance=25.0)
# per-player usage share and threat created (shots/goals via the
# assist-chain link) among throw-ins that travel at least min_distance

long_throw_second_phases(events, min_distance=25.0)
# event-sequence-based flick-on/knockdown detection, restricted to
# those same long throws

Where it’s used: The throw-in section of By set piece – the closest thing throw-ins have to a value model, since there’s no xT path for this type (see By value model).

Penalty placement

What it describes: Where in the goal each penalty was placed – goal_y_norm/ goal_h_norm (0-1 across/up the frame), corner_zone (a 3x3 grid), placement_score (distance from center) – plus per-taker conversion rate.
Requirements: penalty only – penalties are shots, not passes, detected on shot events rather than the pass-based machinery everything else here uses. Pure geometry, no optional dependency needed.

Compute it

from wa_setpieces import penalty_placement_detail, penalty_taker_summary

penalty_placement_detail(events)
penalty_taker_summary(events)

Where it’s used: The penalty section of By set piece; shares its goal-mouth geometry with wa_setpieces.ml.shot_value (By value model).