Original title: SV-WAM: An Efficient Surround-View World-Action Model for End-to-End Autonomous Driving
Jinyang Wang, Shiwei Li, Junjian Wang, Zhiqiang Deng +11 · September 2026
Self-driving cars need to see all around them to turn or change lanes safely. Many systems predict future video to learn how the world works, but generating video takes too much computing power. Because of this, systems often look through only one front camera. To fix this, researchers built SV-WAM. It uses all six cameras around the vehicle. During training, the computer practices predicting future video to learn good driving actions. But a special masking rule stops the driving decisions from looking at the future video tokens. This allows the system to drop the video generator entirely when driving in the real world, running quickly. They also added a rule to keep car corners off road boundaries. It achieved state-of-the-art planning performance and low latency.
Imagine navigating a crowded city with a map app that only shows the street ahead. When you need to change lanes, you can't see what's happening beside you, leading to misses or accidents. In this paper, the system learns from all camera angles during practice drives, but at runtime, it uses a filter that ignores the extra views for decision-making. This keeps it fast while maintaining awareness of the full surroundings during critical moves.
For self-driving cars, missing what's happening behind or beside you during a merge can be dangerous. SV-WAM keeps the full six-camera view while staying fast enough for real-time decisions, improving safety in maneuvers like lane changes.
The abstract does not specify performance in all real-world conditions beyond the tested benchmarks like NAVSIMv2 and nuScenes. It focuses on efficiency and boundary awareness but doesn't address other challenges such as weather variations or unseen scenarios.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
SV-WAM: An Efficient Surround-View World-Action Model for End-to-End Autonomous Driving — Jinyang Wang, Shiwei Li, Junjian Wang et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.