Skip to content

task_metric

compute_q_score(success, now_satisfied_options, initial_satisfied_options)

Partial-success (Q-score) for one episode/env: a full success scores 1.0; otherwise the fraction of goal predicates that were NOT satisfied at episode start but ARE satisfied now, maximized over the alternative goal-state options. Mirrors the pre-refactor inline formula (lives here next to its only caller, TaskMetric). Empty options/no options return 0.0 instead of raising.

Source code in OmniGibson/omnigibson/metrics/task_metric.py 
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
def compute_q_score(
    success: bool,
    now_satisfied_options: Sequence[Sequence[bool]],
    initial_satisfied_options: Sequence[Sequence[bool]],
) -> float:
    """
    Partial-success (Q-score) for one episode/env: a full success scores 1.0; otherwise the fraction
    of goal predicates that were NOT satisfied at episode start but ARE satisfied now, maximized over
    the alternative goal-state options. Mirrors the pre-refactor inline formula (lives here next to its
    only caller, TaskMetric). Empty options/no options return 0.0 instead of raising.
    """
    if success:
        return 1.0
    if not now_satisfied_options:
        return 0.0
    option_scores = []
    for now_opt, init_opt in zip(now_satisfied_options, initial_satisfied_options):
        if len(now_opt) == 0:
            option_scores.append(0.0)
            continue
        newly_satisfied = sum(int((not init) and now) for now, init in zip(now_opt, init_opt))
        option_scores.append(newly_satisfied / len(now_opt))
    return max(option_scores) if option_scores else 0.0