AutoVLA-D: Learn a Trajectory-Geometric Condition from Reasoning for Autonomous Driving VLA
AutoVLA-D: Learn a Trajectory-Geometric Condition from Reasoning for Autonomous Driving VLA
This paper presents AutoVLA-D, a method that introduces TrajCond—a discrete trajectory-geometric condition—to enhance autoregressive VLA planning for autonomous driving. By representing the condition as a 1D-Brohan differential action trajectory and using a two-stage training framework, the approach simultaneously generates discrete and numeric trajectories. Experiments on NAVSIM navtest, nuScenes, and NeuroNCAP show that while discrete trajectories are less effective as direct outputs, they serve as more effective conditions for improving final continuous trajectory prediction in both closed-loop and open-loop driving scenarios. The code is available at https://github.com/ShivaTeam/AutoVLA-D.
The authors introduce TrajCond, a discrete trajectory-geometric condition for autoregressive Vision-Language-Action (VLA) planning in autonomous driving, which they demonstrate improves continuous trajectory prediction on NAVSIM navtest, nuScenes, and NeuroNCAP benchmarks.
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.
- code: Available Artifact link Open source S4
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
rajCond as a 1D-Brohan differential action trajectory and learn it with a two-stage training framework before final continuous trajectory prediction. With the simultaneous autoregressive generation of discrete and numeric trajectories, we name our method AutoVLA-D (Double). Experiments on NAVSIM navtest, nuScenes, and NeuroNCAP show that discrete trajectories are weaker as direct outputs but more effective as TrajCon
Open source S3
arXiv:2611.12382v1 Announce Type: new Abstract: Autoregressive VLA planning usually predicts continuous trajectories as the final output, but driving instructions often describe actions rather than ego states, which weakens direct conditioning on the final continuous trajectory. We address this gap with TrajCond, which introduces a discrete trajectory-geometric condition for VLA planning. Specifically, we represent T
Open source S2
d, consistently improving final continuous trajectory prediction across closed-loop and open-loop driving benchmarks. Code is available at https://github.com/ShivaTeam/AutoVLA-D.
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org