Neuro-Symbolic Drive: Rule-Grounded Faithful Reasoning for Driving VLAs
06:00 · June 24, 2026 · arXiv cs.AI RSS
Driving VLA models incorporating Chain-of-Thought (CoT) reasoning are attractive because they leverage pretrained VLM representations and expose intermediate decisions in natural language, yet current rationales often lack the step-by-step decision semantics needed to keep the rationale causally connected to the planned motion. We introduce Neuro-Symbolic Drive, a neuro-symbolic driving framework that supervises a driving VLA with rule-grounded reasoning traces extracted directly from classical rule-based planners. Our key observation is that rule-based planners are symbolic AI systems that already function as executable reasoning engines: they reason about active safety constraints, search over candidate maneuvers, and select a final trajectory. We instrument these planners in simulation to capture both the executed trajectory and the internal decision trace at each rule-evaluation step. Each trace is serialized into structured rule-grounded reasoning and paired with the trajectory to fine-tune Qwen3.5-4B as a driving VLA. Because these traces are derived directly from the planner states that determine the action, they ensure reasoning is structurally coupled to motion generation by construction, rather than by post-hoc alignment. On our simulator-generated benchmark, detailed rule-grounded reasoning reduces ADE@3s from 0.47 to 0.26 and miss rate from 8.30% to 6.40% under three-camera perception, and from 0.54 to 0.26 and 10.13% to 5.99% under eight-camera perception. Neuro-Symbolic Drive thus converts neuro-symbolic planning logic into structured supervision. Code base: https://github.com/XiangboGaoBarry/Neural-Symbolic-Drive.
Summary
Neuro-Symbolic Drive addresses a core limitation in current Vision-Language-Action models for autonomous driving: while Chain-of-Thought reasoning exposes intermediate decisions in natural language, the generated rationales frequently remain disconnected from the actual motion commands. The framework remedies this by extracting structured reasoning traces directly from classical rule-based planners, which already operate as executable symbolic engines that evaluate safety constraints, enumerate candidate maneuvers, and commit to a final trajectory.
In simulation, the planners are instrumented to record both the executed trajectory and the internal decision states at each rule-evaluation step. These traces are serialized into a unified, rule-grounded reasoning format and paired with the corresponding trajectory data. The resulting dataset is then used to fine-tune Qwen3.5-4B, training the model to generate reasoning that is structurally derived from the same planner states responsible for the motion output. This approach replaces post-hoc alignment with supervision that is causally grounded by construction.
Evaluations on a simulator-generated benchmark show clear gains under both three- and eight-camera perception settings. Average displacement error at three seconds drops from 0.47 m to 0.26 m and from 0.54 m to 0.26 m respectively, while miss rates fall from 8.30 % to 6.40 % and from 10.13 % to 5.99 %. The method therefore converts existing neuro-symbolic planning logic into reliable supervision signals that improve both trajectory accuracy and the behavioral consistency of the accompanying language explanations.
Why it matters
This research is highly relevant for Dutch AI researchers and autonomous system developers because it addresses the critical need for transparent, rule-bound AI in physical environments. Its focus on faithful, explainable reasoning aligns strongly with EU AI Act requirements and the Dutch emphasis on ethical, safe AI deployment.
