This metric evaluates whether a closed-loop rollout actually left its anchor viewpoint and later returned to a previously visited one. It is an official metric for the closed-loop family of the long_sequence dimension. A failed gate no longer blocks subsequent memory metrics: they calculate their own measurements and retain gate_passed=false and gate_failure_reason in their details. Pairs are read from the estimated pose track, not from ground-truth action indices.
Camera-to-world poses {Tt}t=0T−1⊂SE(3) are loaded from the ViPE sidecar beside the prediction. Missing poses make the whole memory track not_applicable. Let pt be translation and Rt rotation. Translation is normalized by the trajectory’s own departure extent E=maxt∥pt−p0∥, then combined with the geodesic rotation angle θ(i,j)=∠(Ri,Rj) (degrees) into a scale-free pose distance:
d(i,j)=E∥pi−pj∥+λ⋅180θ(i,j)(1)
The default is λ=1 (rotation_weight). When E is degenerate (an in-place rotation loop), the translation term is dropped and the distance falls back to rotation only.
Departure requires the farthest pose from the anchor to clear several pair-thresholds:
ddep=tmaxd(0,t)≥α⋅τ,α=3,τ=0.15(2)
i.e. min_departure_factor * max_pair_distance. Return then, for each frame j, selects the geometrically closest earlier frame i⋆=argmini≤j−gd(i,j) at least Δtmin seconds away (default 1 s, g=⌈Δtmin⋅fps⌉). The pair is kept if d(i⋆,j)≤τ and the intervening excursion satisfies
e(i⋆,j)=i⋆<k<jmaxd(i⋆,k)≥β⋅τ,β=3(3)
(min_excursion_factor * max_pair_distance). The excursion test stops slow continuous motion from pairing every frame with its own recent past. A model that never walks back is reported as having no revisit, not as having forgotten the scene.
Sgate∈{0,1}. In WorldAtlas this algorithm is implemented as the official metric memory_return_gate (pose_loop_gate backend). Under the default normalization config, raw and normalized outputs lie in {0,1}. Report this official metric as a pass rate alongside the other memory scores; aggregate weights follow the published leaderboard rule. When strict revisit pairs are absent, scoring uses nearest earlier poses satisfying the same minimum time gap, without the return-distance or excursion thresholds (pair_source=nearest_pose_fallback). Scene F1 uses a supported revisit episode when available; after a failed gate with no supported episode, it compares the first and second halves of the pose track (split_source=temporal_midpoint_fallback), requiring at least two pose frames per half. These fallback scores can reflect viewpoint/control differences and should be read together with the gate. Missing inputs, unavailable backends, insufficient pair support or unreliable half-life fits still produce an explicit unavailable result.
