Closed-loop planning safety of refined forecasts

Determine whether forecasts refined by FutureLift translate into safer planning decisions under closed-loop interaction.

Background

FutureLift is evaluated on offline motion-forecasting benchmarks using high-definition map context. Although the method improves trajectory-based forecasting metrics, the paper does not establish that these geometric improvements improve downstream autonomous-vehicle planning or safety when predictions and actions interact over time.

The authors identify closed-loop evaluation as a necessary extension and specifically call for assessing whether FutureLift-refined forecasts produce safer decisions under interactive planning, rather than merely better open-loop displacement metrics.