- The paper introduces a neuro-symbolic framework that leverages rule-based planner traces to provide causally faithful, action-bearing reasoning in driving VLAs.
- It employs scenario-aware teacher selection and a unified four-slot schema to significantly reduce trajectory errors in simulation scenarios.
- The approach enhances auditability and interpretability by coupling trajectory predictions with language rationales that mirror underlying decision semantics.
Neuro-Symbolic Drive: Rule-Grounded Faithful Reasoning for Driving VLAs
Motivation and Background
The paper addresses a persistent gap in vision-language autonomous driving agents (VLAs): the disconnect between natural language rationales and executable actions. Driven by advances in Chain-of-Thought (CoT) prompting, many recent driving VLAs leverage language explanations to increase interpretability. However, prior approaches largely rely on post-hoc or free-form rationales, which are only weakly or non-causally correlated with the agent's actual motion planning. This supervision mismatch results in unfaithful reasoning traces, undermining auditability and safety-critical assurance in autonomous driving.
Contrary to imitation-based VLA supervision, classical rule-based planners execute explicit safety checks, candidate maneuvers, and structured trajectory selection procedures. Their decision traces are tightly coupled to the selected motion and thus inherently action-bearing. The central thesis of this work is to repurpose these planner execution traces as structured supervision for driving VLAs, ensuring semantic coupling and interpretability by construction rather than alignment.
Neuro-Symbolic Drive Framework
The NEURO-SYMBOLIC DRIVE framework operates in simulation by instrumentation of three complementary rule-based planner teachers: IDM (Intelligent Driver Model), IDM-MOBIL (which incorporates lane-change decision logic), and PDM-Closed (proposal-based closed-loop planner). Each planner’s execution generates both a trajectory and an internal reasoning trace at each timestep. A lightweight abstraction serializes heterogeneous traces from the different planners into a unified four-slot reasoning schema: scene context, active constraint, candidate maneuver evaluation, and decision outcome.
For each simulated scenario, a scenario-aware selector assigns supervision responsibility to the most capable planner (based on closed-loop quality metrics such as CCLS-R). This ensures that reasoning traces are context-appropriate and robust, avoiding generalization errors that arise from assigning a single planner to all scenarios.
The driving VLA backbone is Qwen3.5-4B, fine-tuned end-to-end with concatenated rule-grounded reasoning sequences and corresponding future trajectories.
Supervision Design and Variants
The paper proposes three supervision modalities:
- Trajectory-only (no reasoning): The model receives only the future trajectory, with no intermediate reasoning.
- Concise rule-grounded reasoning: Each supervision trace exposes the key constraints, tested candidates, and maneuver decisions in a compact, quantitative format.
- Detailed rule-grounded reasoning: Traces are elaborated with step-wise decision semantics, narrative justifications, and rejection rationales for candidate maneuvers.
Both reasoning variants share the four-slot schema but differ in token length and decision trace granularity.
Empirical Evaluation
Experiments are conducted on the NSD-SIM dataset, covering 520 scenarios and 77k+ timestep examples, over two camera configurations (three-view and eight-view). Metrics evaluated include Average Displacement Error (ADE), Final Displacement Error (FDE), Average/Final Heading Error (AHE/FHE), and Miss Rate (MR) at 3 seconds.
Key numerical results:
- Three-camera setting: Detailed reasoning reduces ADE@3s from 0.47 (no reasoning) to 0.26 and MR@3s from 8.30% to 6.40%.
- Eight-camera setting: Detailed reasoning reduces ADE@3s from 0.54 to 0.26 and MR@3s from 10.13% to 5.99%.
- Planner-source ablation: Scenario-aware multi-teacher selection consistently outperforms single-teacher supervision. PDM-Closed dominates intersection/turn scenarios; IDM-MOBIL is superior for lateral maneuvers and agent interactions; IDM excels in conservative crawl/yield contexts.
- Reasoning variant comparison: Concise reasoning achieves most of the gain; detailed reasoning provides additional improvement on long-horizon and heading metrics, at the cost of increased sequence length and inference latency.
These results clearly establish that the performance improvements are driven by rule-grounded reasoning, not sensor suite expansion or simple data augmentation.
Theoretical and Practical Implications
The framework eliminates the supervision mismatch by synchronizing reasoning traces with trajectory supervision from the same execution. This design strictly enforces action-bearing rationales in the VLA output, fundamentally improving the fidelity of language explanations in driving agents.
Practically, NEURO-SYMBOLIC DRIVE increases the inspectability and auditability of autonomous driving models. Each predicted trajectory is accompanied by a causally faithful language rationale, traceable to programmatic decision semantics. This enables advanced debugging, regulatory review, and hybrid formal verification workflows.
Theoretically, the method extends neuro-symbolic integration in AI for autonomous driving, demonstrating that executable planner traces are potent sources of structured supervision. The principle is generalizable: any expert system with an internal decision trace (constraint solvers, rule engines, etc.) is a candidate supervision source for action-faithful reasoning in agentic models.
Limitations and Future Directions
All experiments are simulation-based. Real-world closed-loop deployment and generalization remain open challenges due to the simulation-to-real gap and planner deficiencies. Reasoning depth increases token and compute cost; concise traces are preferable for deployment. The VLA inherits the capability boundaries of the underlying planner teachers. Addressing planner blind spots, expanding teacher diversity, and formalizing reasoning trace extraction in real fleet data are promising future directions.
Direct integration of NEURO-SYMBOLIC DRIVE into formal verification pipelines, reinforcement fine-tuning for scenario-specific reward modeling, and extension to agentic collaborative driving systems are warranted.
Conclusion
NEURO-SYMBOLIC DRIVE presents a robust neuro-symbolic supervision pipeline for vision-language driving agents, leveraging executable rule-based planner traces for causally faithful reasoning. Empirical results verify substantial gains in trajectory alignment and behavioral consistency across a broad range of driving scenarios. Scenario-aware teacher selection and structured reasoning abstraction establish new standards for interpretable and action-bearing supervision in autonomous driving VLAs. The methodology is broadly extensible to other agentic AI domains with structured expert decision traces (2606.23938).