Accurate Forecasts Do Not Ensure Safe Decision

Jaehyun Pyun, Seunghun Moon, Suk-Ju Kang

Keywords: Trajectory Forecasting, Risk-Aware Planning, Benchmarking

Abstract: Pedestrian forecasting is still largely evaluated with retrospective best-of-$K$ metrics such as minADE and minFDE, although planners must commit to safe actions before the future is observed. We introduce the Trajectory-Decision benchmark protocol (TrajD), a compact decision-centric stress test built on the TrajImpute missing-trajectory benchmark dataset. TrajD fixes the incomplete-history carrier, ego-action set, planner, ego-agent selection rule, and ground-truth decision metrics, then audits five stochastic sources--MoFlow, SingularTrajectory, TUTR, NMRF, and SocialGAN--through one forecast-to-decision pipeline. Source planning leaves substantial avoidable collision risk. As a reference action-risk alignment method, a lightweight two-fold out-of-fold (OOF) aligner leaves source samples unchanged but rescales planner action scores, reducing collision by about one third on average; a less conservative operating point partially recovers goal progress. Results show that forecasting benchmarks should report downstream safe decisions alongside geometric sample accuracy.

Submission Number: 55