GatedVLA with Camera-Based Emergency Braking: AlpaSim E2E Closed Loop Challenge Technical Report
GatedVLA with Camera-Based Emergency Braking: AlpaSim E2E Closed Loop Challenge Technical Report
This technical report introduces GatedVLA, a vision-language-action policy designed for the Physical AI AV track. The system pairs a driving-pretrained Qwen-Drive-1.0-4B model with a flow-matching action expert enhanced by a soft visual mask. In local closed-loop evaluations across 150 scenes, the masked policy improved the mean scene score from 0.479 to 0.517. When augmented with a rule-based automatic emergency braking module, the mean score further increased to 0.599, and the mean number of scenes with at-fault collisions decreased from 34.0 to 11.0.
The report details a technical entry for the AlpaSim E2E Closed Loop Challenge, combining a vision-language-action policy with rule-based automatic emergency braking to improve scene scores and reduce collision rates in simulated driving scenarios.
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
scene score
0.479–0.517
View original evidence
cross-attention. Compared to an identically trained baseline, the masked policy achieves comparable overall open-loop error while reducing error on left turns. In local closed-loop evaluation across the same 150 scenes (two rollouts per scene), the visual mask increases the mean scene score from 0.479 to 0.517 (paired difference +0.038, 95% bootstrap CI [-0.018, +0.093], wins/losses 37/26, sign test p = 0.21). Incorp
Open source S3mean scene score increase
0.038
View original evidence
cross-attention. Compared to an identically trained baseline, the masked policy achieves comparable overall open-loop error while reducing error on left turns. In local closed-loop evaluation across the same 150 scenes (two rollouts per scene), the visual mask increases the mean scene score from 0.479 to 0.517 (paired difference +0.038, 95% bootstrap CI [-0.018, +0.093], wins/losses 37/26, sign test p = 0.21). Incorp
Open source S3scene score with AEB
0.599
View original evidence
orating a rule-based automatic emergency braking (AEB) module at deployment further increases the mean score to 0.599 (+0.082, CI [+0.023, +0.143]) and reduces the mean number of scenes with at-fault collisions from 34.0 to 11.0.
Open source S4
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
cross-attention. Compared to an identically trained baseline, the masked policy achieves comparable overall open-loop error while reducing error on left turns. In local closed-loop evaluation across the same 150 scenes (two rollouts per scene), the visual mask increases the mean scene score from 0.479 to 0.517 (paired difference +0.038, 95% bootstrap CI [-0.018, +0.093], wins/losses 37/26, sign test p = 0.21). Incorp
Open source S3
orating a rule-based automatic emergency braking (AEB) module at deployment further increases the mean score to 0.599 (+0.082, CI [+0.023, +0.143]) and reduces the mean number of scenes with at-fault collisions from 34.0 to 11.0.
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org