Multimodal trajectory prediction can broaden behavioral coverage in end‑to‑end autonomous driving, yet current approaches are constrained by sparse scene representations. Incomplete evidence leads to low‑quality candidate generation and unreliable ranking among geometrically similar trajectories. On a register‑based baseline, poor candidates account for 19.74% of the set, and the oracle‑best candidate ranks only 33.9th on average.
We introduce RRDrive, which leverages risk‑aware occupancy as a dense, temporally aligned, trajectory‑queryable representation. Its global structure guides high‑quality multimodal generation, while candidate‑conditioned risk queries enable fine‑grained selection. We also build the RiskOcc4D‑NAVSIM dataset with automatically annotated risk labels.
Results show RRDrive achieves a selected‑trajectory PDMS of 0.951, a 1.5% relative gain over the baseline (0.937), and improves average candidate PDMS by 7.7%. In challenging scenes, candidate PDMS rises by 30.2% and the Spearman correlation among good candidates increases from 0.26 to 0.67 (+0.41).
To move beyond the oracle setting, we develop an external RiskOcc predictor—a perception module that directly estimates risk‑aware occupancy from sensor inputs. Its competitive performance confirms the feasibility of the representation.
Review