Driving · 2026-11-04
Trajectory Evaluation via Rollout-Free World Model for End-to-end Autonomous Driving
Junzhe Wang, Jinqing Zhang, Qingjie Liu
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
RFWorld proposes rollout-free world-model evaluation for end-to-end autonomous driving. Instead of predicting candidate-conditioned future scenes (rollout) for each candidate, it builds a shared scene memory from the current BEV and shapes it via lightweight training-time temporal readout with semantic supervision. At inference, candidate trajectories score by sparse spatial readout from the shared memory, reducing computation and avoiding rollout error accumulation.
Why it matters
Candidate trajectory evaluation needs future-awareness but rollout is expensive and error-prone. RFWorld reframes the world model as a planning-oriented shared memory plus candidate queries, enabling rollout-free scoring while keeping training overhead low. The authors report state-of-the-art performance on NAVSIM and improved results on Bench2Drive.
Method
- Build a trajectory-agnostic shared scene memory from the current BEV (trajectory anchors interact with current scene to produce candidates; shared memory is reused across training and inference).
- Shape the shared memory during training using lightweight temporal readout with semantic supervision at multiple future timestamps.
- At inference, score candidates by sparse spatial readout via deformable attention from the shared memory (no explicit future semantic decoding and no candidate-wise rollout).
Abstract (from arXiv)
Although existing candidate trajectory evaluators have substantially improved end-to-end planning accuracy, their underlying mechanism remains suboptimal. Model-based evaluators typically make scoring future-aware by explicitly predicting candidate-conditioned future states over the planning horizon, which we term rollout. However, rollout incurs high inference cost and accumulates prediction errors. We argue that planning does not require reconstructing a full future scene for each candidate, but only future-relevant information for reliable comparison. We therefore propose RFWorld, a Rollout-Free World model for end-to-end trajectory evaluation. RFWorld constructs a shared scene memory and shapes it into a future-readable representation via lightweight training-time temporal readout with semantic supervision, keeping training overhead low. At inference time, candidate trajectories directly query the shared memory via sparse spatial readout for scoring, reducing computation by roughly the planning horizon times the number of candidate trajectories, while avoiding recursive rollout errors. Experiments on NAVSIM and Bench2Drive show that RFWorld learns a more planning-aligned world representation and achieves state-of-the-art performance.
Related papers
- PlanWAM: Planning-Shaped Future Representations for End-to-End Autonomous Driving
- Latent-WAM++: Shaping World and Action Representations for End-to-End Autonomous Driving
- DriftingPlanner: Drifting Generative Models for One-Step Trajectory Planning in End-to-End Autonomous Driving
- Directional Pullback Regularization for Long-Horizon Rollout of Latent World Models in Autonomous Driving
- World4Scorer: Outcome-Grounded World Modeling for Autonomous Driving
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-11-03 · Indexed by Pub-AI: 2026-11-04 UTC
Trajectory Evaluation via Rollout-Free World Model for End-to-end Autonomous Driving
Preprint posted: 2026-11-03 · Original record ↗
- arxiv
- 2611.05072
Explicit source versions (1)
- v1 · 2026-11-03
- Preprint available
Source date: 2026-11-03 · First detected: 2026-11-04 UTC · Historical baseline
Source evidence (2)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2611.05072v1; arxiv:announce_type=new; item pubDate (official metadata)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Although existing candidate trajectory evaluators have substantially improved end-to-end planning accuracy, their underlying mechanism remains suboptimal. Model-based evaluators typically make scoring future-aware by explicitly predicting candidate-conditioned future states over the planning horizon, which we term rollout. However, rollout incurs high inference cost and accumulates prediction errors. We argue that planning does not require reconstructing a full future scene for each candidate, but only future-relevant information for reliable comparison. We therefore propose RFWorld, a Rollout-Free World model for end-to-end trajectory evaluation. RFWorld constructs a shared scene memory and shapes it into a future-readable representation via lightweight training-time temporal readout with semantic supervision, keeping training overhead low. At inference time, candidate trajectories directly query the shared memory via sparse spatial readout for scoring, reducing computation by roughly the planning horizon times the number of candidate trajectories, while avoiding recursive rollout errors. Experiments on NAVSIM and Bench2Drive show that RFWorld learns a more planning-aligned world representation and achieves state-of-the-art performance.
