Temporal Metrics¶

Use temporal metrics when error accumulation across rollout time matters more than a single final scalar.

Reference¶

RoboMetrics internal heuristic for temporal error growth.

Quick Example¶

import numpy as np
from robometrics import compounding_error_index

errors = np.array([0.2, 0.25, 0.4, 0.8])
score = compounding_error_index(errors)
print(score)

Metrics¶

temporal_drift(predicted, reference) -> float¶

Formula: least-squares slope of per-timestep L2 error over timestep index. Reference: RoboMetrics internal temporal drift diagnostic Unit: error/timestep Direction: lower is better

Temporal drift reports whether rollout error trends upward or downward over time. Negative values mean error decreases over the rollout.

action_jerk(actions, dt) -> float¶

Formula: mean squared L2 norm of the second finite difference of actions. Reference: RoboMetrics internal control smoothness diagnostic Unit: action^2/s^4 Direction: lower is better

Action jerk is a scalar control-sequence smoothness penalty.

control_smoothness(actions, dt) -> float¶

Formula: 1 / (1 + action_jerk(actions, dt)). Reference: RoboMetrics internal control smoothness score Unit: score Direction: higher is better

Control smoothness maps action jerk into a bounded score.

long_horizon_drift(predicted, reference) -> float¶

Formula: later-weighted mean L2 rollout error with weights 1..T. Reference: RoboMetrics internal long-horizon drift diagnostic Unit: state units Direction: lower is better

Long-horizon drift penalizes the same error more when it appears later in a rollout.

compounding_error_index(errors) -> float¶

Formula: final error divided by initial finite error, with stable handling near zero. Reference: RoboMetrics internal heuristic Unit: ratio Direction: lower is better

Compounding error index highlights whether a rollout gets worse as prediction or control steps compound.