Why

alpamayo_r1.metrics.distance_metrics already implements ADE / minADE / corner_distance, but the standard companion metric for trajectory prediction — Final Displacement Error — is missing.

ADE measures average L2 error along the trajectory; FDE measures L2 error at the last timestep. Reporting one without the other distorts the picture in two opposite ways:

• A model that drifts early but recovers looks worse under ADE than under FDE.
• A model that nails the early steps but veers off at the end looks better under ADE than under FDE.

Every major trajectory-prediction benchmark (Waymo Open Motion, nuScenes, Argoverse) reports both. Adding FDE here closes a real evaluation gap and aligns Alpamayo's metric outputs with the literature.

What

Two new functions in src/alpamayo_r1/metrics/distance_metrics.py that mirror compute_ade / compute_minade exactly — same shapes, same kwargs, same error semantics.

compute_fde(pred_xyz, gt_xyz, timestep_horizon=None, only_xy=True) -> torch.Tensor

Returns [B, N, K]. timestep_horizon=H evaluates FDE at the H-th timestep (1-indexed) instead of the final one — useful for reporting FDE@3s. Raises ValueError on timestep_horizon > T (mirroring compute_ade).

compute_minfde(pred_xyz, gt_xyz, disable_summary=False, timestep_horizons=[5, 10, 30, 50], only_xy=True, time_step=0.1) -> dict[str, Tensor]

Returns per-batch tensors of shape [B]:

• min_fde: K-min FDE at the final timestep, averaged over the N groups.
• min_fde/by_t={H:.1f}: K-min FDE at the timestep int(H / time_step), for each valid t in timestep_horizons. Out-of-range horizons are silently dropped (mirrors compute_minade).
• <key>_std: stdev across the N groups, added by summarize_metric when N > 1 and disable_summary=False.

Key naming is parallel to compute_minade so dashboards / loggers that already group on min_ade* automatically pick up min_fde*.

Tests

src/alpamayo_r1/metrics/test_fde.py — 11 pytest cases, no GPU / no HF auth:

• compute_fde shape, non-negativity, manual final-step L2 equivalence.
• only_xy=False uses the 3D vector and returns a value >= the XY-projection FDE.
• timestep_horizon=H selects index H-1.
• timestep_horizon > T raises ValueError.
• FDE is independent of intermediate timesteps (perturbing them by +100 leaves FDE unchanged).
• FDE differs from ADE for non-trivial trajectories (sanity).
• compute_minfde emits all documented keys + _std variants when N > 1.
• Out-of-range horizons in compute_minfde are silently skipped.
• disable_summary=True drops _std.
• min_fde equals compute_fde(...).min(K).mean(N).

Local verification

Validated against torch (no GPU/HF needed) by exec'ing the new functions in isolation — all 6 core invariants pass. Output:

PASS: compute_fde shape (torch.Size([2, 3, 4])) and non-negative
PASS: compute_fde matches manual L2 at final step
PASS: compute_fde ignores intermediate timesteps
PASS: compute_minfde emits all documented keys + _std variants
PASS: compute_minfde silently skips out-of-range horizons
PASS: compute_fde rejects horizon > T

Migration

None. This is purely additive — existing call sites that import compute_ade / compute_minade are untouched.