Neuro-Symbolic Drive: rule-grounded reasoning cuts driving errors by 45%

Driving vision-language-action (VLA) models often use Chain-of-Thought (CoT) reasoning to explain their decisions, but these rationales frequently lack the step-by-step semantic link to actual motion planning. Neuro-Symbolic Drive solves this by extracting decision traces from classical rule-based planners—symbolic AI systems that already function as executable reasoning engines. These planners evaluate safety constraints, search maneuvers, and select trajectories, producing a structured trace at each step. The framework serializes these traces into rule-grounded reasoning chains and pairs them with the executed trajectory to fine-tune a Qwen3.5-4B model as a driving VLA. Because the supervision comes directly from the planner states that determine the action, the reasoning is inherently coupled to motion generation, not post-hoc aligned.

On a simulator-generated benchmark, Neuro-Symbolic Drive delivered significant improvements across camera configurations. Under three-camera perception, Average Displacement Error at 3 seconds (ADE@3s) dropped from 0.47 to 0.26 (a 45% reduction), and miss rate fell from 8.30% to 6.40%. With eight-camera perception, ADE@3s improved from 0.54 to 0.26 (52% reduction) and miss rate from 10.13% to 5.99%. These results show that grounding neural reasoning in symbolic planning logic not only makes the rationale faithful but also significantly enhances driving performance, paving the way for more interpretable and reliable autonomous driving systems.