Research map / Label-efficient and robust learning / Recognition and long-tail
Trajectory Evaluation via Rollout-Free World Model for End-to-end Autonomous Driving
Junzhe Wang, Jinqing Zhang, Qingjie LIU
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a rollout-free world-model trajectory evaluator for end-to-end autonomous driving that scores candidate trajectories without explicitly predicting candidate-conditioned future states.
Problem
Model-based candidate trajectory evaluators in end-to-end driving make scoring future-aware by explicitly predicting candidate-conditioned future states (rollout), which incurs high inference cost and accumulates prediction errors.
Approach
Replaces explicit candidate-conditioned rollout with a rollout-free world model that makes trajectory scoring future-aware implicitly, avoiding the cost and error accumulation of predicting future states for each candidate.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add world model branch without future supervision (ablation) | RFWorld · NAVSIM-v1 | PDMS: 88.9 -> 88.4 | Hurt |
| add explicit future supervision (no world model transformer) (ablation) | RFWorld · NAVSIM-v1 | PDMS: 88.9 -> 89.4 | Improved |
| add world model on top of future supervision (full RFWorld) (ablation) | RFWorld · NAVSIM-v1 | PDMS: 89.4 -> 90.5 | Improved |
| supervise more future timestamps (0 to 8) (ablation) | RFWorld · NAVSIM-v1 | PDMS: 88.4 -> 90.5 | Improved |
| increase number of test-time candidate anchors from 256 to 1024 (ablation) | RFWorld · NAVSIM-v1 | PDMS: 89.6 -> 90.5 | Improved |
| increase test-time candidate anchors from 256 to 1024 (ablation) | WoTE · NAVSIM-v1 | PDMS: 88.3 -> 88.7 | Improved |
| replace per-candidate rollout with shared-memory sparse readout | RFWorld vs WoTE · NAVSIM-v1 | inference latency (ms): 49.2 -> 37.5 | Improved |
| replace per-candidate rollout with shared-memory readout | RFWorld vs WoTE · NAVSIM-v1 | training GPU memory (GB): 33.9 -> 8.7 | Improved |
| use rollout-free world model scoring instead of WoTE | RFWorld vs WoTE · NAVSIM-v1 | PDMS: 88.3 -> 90.5 | Improved |
| evaluate in closed loop on Bench2Drive | RFWorld vs TCP · Bench2Drive | Driving Score: 59.90 -> 62.38 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- ResWorld — ResWorld: Temporal Residual World Model for End-to-End Autonomous Driving (ICLR 2026) · better on 1 of 1 shared result
- DriveSuprim — DriveSuprim: Towards Precise Trajectory Selection for End-to-End Planning (AAAI 2026) · better on 1 of 1 shared result
- Transfuser — TransFuser: Imitation With Transformer-Based Sensor Fusion for Autonomous Driving (TPAMI 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation (NeurIPS 2026 · similarity 0.88)
- Future-Aware End-to-End Driving: Bidirectional Modeling of Trajectory Planning and Scene Evolution (NeurIPS 2025 · similarity 0.88)
- ChainFlow-VLA: Causal Flow Planning with Vision-Language Models (NeurIPS 2026 · similarity 0.87)
- Driving in the Occupancy World: Vision-Centric 4D Occupancy Forecasting and Planning via World Models for Autonomous Driving (AAAI 2025 · similarity 0.87)
- DriveFuture: Future-Aware Latent World Models for Autonomous Driving (NeurIPS 2026 · similarity 0.87)
- Generative Scenario Rollouts for End-to-End Autonomous Driving (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
