Neuro-Symbolic Drive: Rule-Grounded Reasoning for Driving VLAs

Neuro-Symbolic Drive, presented in an arXiv paper submitted on 22 Jun 2026 by Xiangbo Gao et al., supervises a driving VLA by pairing trajectories with rule-grounded reasoning traces extracted directly from classical rule-based planners. The method fine-tunes Qwen3.5-4B using those serialized traces and yields concrete reductions in ADE@3s and miss rate on a simulator-generated benchmark.

How does Neuro-Symbolic Drive work?

It instruments classical rule-based planners in simulation to capture both the executed trajectory and the internal decision trace at each rule-evaluation step, serializes each trace into structured rule-grounded reasoning, and pairs that reasoning with the trajectory to fine-tune Qwen3.5-4B. The authors observe that rule-based planners already function as executable reasoning engines: they evaluate active safety constraints, search candidate maneuvers, and select a final trajectory. By extracting the planners internal states and decisions, the team generates supervision that ties natural-language reasoning traces directly to the motion the planner produced, rather than aligning explanations to motion after the fact.

What improvements did the rule-grounded reasoning deliver?

On the simulator-generated benchmark, adding detailed rule-grounded reasoning reduced ADE@3s from 0.47 to 0.26 and cut miss rate from 8.30% to 6.40% under three-camera perception. Under eight-camera perception, ADE@3s fell from 0.54 to 0.26 and miss rate dropped from 10.13% to 5.99%. The paper frames these results as evidence that traces derived directly from planner states keep the reasoning structurally coupled to motion generation by construction. The implementation fine-tunes Qwen3.5-4B as the driving VLA and pairs the serialized rule evaluations with the corresponding trajectories as supervision. The authors also link to a code base in the paper text: "this https URL".

Why it matters

Neuro-Symbolic Drive addresses a persistent gap in driving VLAs: rationales that look plausible in language but are not causally connected to the actions they claim to explain. By instrumenting symbolic planners and using their actual decision traces as supervision, the approach forces the VLAs intermediate reasoning to reflect the same steps that produced the motion. That structural coupling can increase faithfulness of explanations and make the models step-by-step decisions actionable for downstream safety or verification workflows.

What to watch

Confirmatory signals would be reproducing the reported ADE@3s and miss-rate reductions on held-out simulation scenarios and on real-world datasets, and adapting the rule-grounded supervision to other pretrained VLMs beyond Qwen3.5-4B. The paper includes a link to a code base ("this https URL") that will be the first place to check for replication materials and further experiments.