Neuro-Symbolic AI Improves Autonomous Driving Reasoning

Driving Visual Language-Action (VLA) models that incorporate Chain-of-Thought (CoT) reasoning are appealing because they leverage pre-trained VLM representations and articulate intermediate decisions in natural language. However, current rationales often lack the precise step-by-step decision semantics needed to ensure a causal link between the reasoning and the actual planned motion. To address this, the Neuro-Symbolic Drive framework has been introduced. This neuro-symbolic approach supervises a driving VLA with reasoning traces that are directly grounded in rules and extracted from classical rule-based planners. The core insight is that rule-based planners inherently function as executable reasoning engines, capable of evaluating safety constraints, exploring maneuvers, and selecting trajectories. By instrumenting these planners in simulation, the researchers captured both the executed trajectory and the internal decision trace at each rule-evaluation step. These traces, serialized into structured rule-grounded reasoning, are then paired with the trajectory to fine-tune a VLA model like Qwen3.5-4B. This method ensures that the reasoning is structurally coupled to motion generation by design, rather than through post-hoc alignment. Empirical results on a simulator-generated benchmark showed significant reductions in driving errors, including a decrease in ADE@3s from 0.47 to 0.26 and miss rate from 8.30% to 6.40% under three-camera perception.