Researchers have shown that an end‑to‑end driving system can evaluate candidate future paths without repeatedly predicting whole future scenes. RFWorld builds a shared scene memory from a bird‑eye view (BEV) representation. During training the model learns a future‑readable memory by reading the BEV sequence over time while being supervised with semantic labels. At run‑time each candidate trajectory queries this memory through a sparse spatial readout to get a reward score. The approach removes the need to rollout future states for every trajectory and cuts inference cost roughly by the planning horizon multiplied by the number of candidates.

On the NAVSIM‑v1 benchmark, RFWorld achieved a Predictive Driver Model Score (PDMS) of 90.5 using a ResNet‑34 backbone, surpassing all other state‑of‑the‑art end‑to‑end methods evaluated with the same backbone. The method also showed stronger alignment with human‑driven comfort, safety, and compliance metrics.

The authors note that RFWorld’s transformer‑based latent world model struggles with very high‑resolution future predictions on dense BEV features, and that its performance depends on the quality of semantic supervision used during training. In addition, the approach can err if the underlying perception backbone delivers inaccurate features for road geometry or distant traffic participants, which might lead to drivable‑area exits or collisions.

Overall, the study demonstrates that sharing a future‑aware memory can improve trajectory evaluation efficiency and accuracy, but also highlights that further work is needed to extend the method to higher‑resolution predictions and to reduce dependence on labeled semantic data.