Scene-Adaptive Readout and Decoding in Diffusion Driving VLAs
Scene-Adaptive Readout and Decoding in Diffusion Driving VLAs
This research paper introduces a scene-adaptive readout and decoding framework for diffusion driving vision-language-action models (VLAs). By analyzing trajectory readout across network depth, the authors identify that late-layer representations support useful predictions after a single forward sweep while early layers benefit from continued decoding. The proposed method uses a scene-conditioned readout Mixture of Experts (MoE) to adaptively combine predictions and applies additional refinement only when necessary. Evaluated on Waymo E2E and nuScenes benchmarks, the framework improves validation Relative Frequency of Success (RFS) by up to 0.56, achieves an end-to-end speedup of up to 6.0x on Waymo E2E, and consistently reduces trajectory error on nuScenes. The code is available at https://github.com/sjyu001/SARD.
The authors propose a scene-adaptive readout and decoding framework for diffusion driving vision-language-action models, demonstrating improved planning quality and inference speedup on Waymo E2E and nuScenes benchmarks without independent verification of deployment readiness.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
speedup factor
6
View original evidence
by up to 0.56 while also improving test performance, and achieves an end-to-end speedup of up to $6.0\times$ on Waymo E2E, while consistently reducing trajectory error on nuScenes. These results demonstrate that scene-level readout and decoding allocation improves planning quality while avoiding unnecessary decoding. Our code is available at https://github.com/sjyu001/SARD
Open source S5RFS improvement
≤0.56
View original evidence
by up to 0.56 while also improving test performance, and achieves an end-to-end speedup of up to $6.0\times$ on Waymo E2E, while consistently reducing trajectory error on nuScenes. These results demonstrate that scene-level readout and decoding allocation improves planning quality while avoiding unnecessary decoding. Our code is available at https://github.com/sjyu001/SARD
Open source S5
- code: Available Artifact link Open source S5
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
by up to 0.56 while also improving test performance, and achieves an end-to-end speedup of up to $6.0\times$ on Waymo E2E, while consistently reducing trajectory error on nuScenes. These results demonstrate that scene-level readout and decoding allocation improves planning quality while avoiding unnecessary decoding. Our code is available at https://github.com/sjyu001/SARD
Open source S5
ng framework. A scene-conditioned readout MoE adaptively weights and combines trajectory predictions from layer-specific late-layer representations after a single backbone sweep, while an adaptive decoding policy applies additional native refinement only when required by the scene. We evaluate our framework on two diffusion driving VLAs across the Waymo E2E and nuScenes benchmarks. Our method improves validation RFS
Open source S4
Scene-Adaptive Readout and Decoding in Diffusion Driving VLAs
Open source S1
arXiv:2611.02076v1 Announce Type: cross Abstract: Diffusion driving vision-language-action models repeatedly execute the backbone at every denoising stage, incurring substantial inference cost, while it remains unclear whether all scenes require the same decoding computation. To distinguish when a trajectory becomes readable from how much further refinement it requires, we analyze trajectory readout across network de
Open source S2
pth and decoding progress. We identify a depth-decoding dissociation in which late-layer representations already support useful trajectory readout after the first forward sweep, whereas early-layer representations benefit more from continued decoding. We further find that both the preferred readout layer and decoding depth vary across scenes. Based on these observations, we propose a scene-adaptive readout and decodi
Open source S3
Source：arXiv Robotics — research abstracts · arxiv.org