A Neuro-Symbolic Framework Combining Inductive and Deductive Reasoning for Autonomous Driving Planning

Hongyan Wei; Wael AbdAlmageed

cs.CVMar 12, 2026

A Neuro-Symbolic Framework Combining Inductive and Deductive Reasoning for Autonomous Driving Planning

Authors: Hongyan Wei, Wael AbdAlmageed

Organizations: Clemson University

Abstract

Existing end-to-end autonomous driving models rely heavily on purely data-driven inductive reasoning. This "black-box" nature leads to a lack of interpretability and absolute safety guarantees in complex, long-tail scenarios. To overcome this bottleneck, we propose a novel neuro-symbolic trajectory planning framework that seamlessly integrates rigorous deductive reasoning into end-to-end neural networks. Specifically, our framework utilizes a Large Language Model (LLM) to dynamically extract scene rules and employs an Answer Set Programming (ASP) solver for deterministic logical arbitration, generating safe and traceable discrete driving decisions. To bridge the gap between discrete symbols and continuous trajectories, we introduce a decision-conditioned decoding mechanism that transforms high-level logical decisions into learnable embedding vectors, simultaneously constraining the planning query and the physical initial velocity of a differentiable Kinematic Bicycle Model (KBM). By combining KBM-generated physical baseline trajectories with neural residual corrections, our approach inherently guarantees kinematic feasibility while ensuring a high degree of transparency. On the nuScenes benchmark, our method comprehensively outperforms the state-of-the-art baseline MomAD, reducing the L2 mean error to 0.57 m, decreasing the collision rate to 0.075%, and optimizing trajectory prediction consistency (TPC) to 0.47 m.

Figures & tables

Figure 1: Overall framework. The inductive stream (neural) and deductive stream (symbolic) run in parallel. The ASP engine produces discrete driving decisions which are conditioned into the Physics-Constrained Motion Decoder via dual paths: Path 1 modulates planning queries

\mathbf{Q}^{\prime}_{\text{plan}}=\mathbf{Q}_{\text{plan}}\oplus\mathbf{d}

(spatial shape), and Path 2 biases the KBM initial velocity

v_{0}^{\prime}=v_{0}+b_{v}

(speed profile).

Figure 2: Detailed pipeline of the deductive reasoning engine. The LLM extracts semantic rules from perception facts via prompts, which are then combined with a static logical base and evaluated by the ASP solver (Clingo) to deduce final deterministic driving decisions.

Method	Input	Backbone	$L_{2}$ (m) $\downarrow$				Col. Rate (%) $\downarrow$				TPC (m) $\downarrow$				FPS $\uparrow$
Method	Input	Backbone	1s	2s	3s	Avg.	1s	2s	3s	Avg.	1s	2s	3s	Avg.	FPS $\uparrow$
UniAD † [ 13 ]	Camera	ResNet101	0.48	0.96	1.65	1.03	0.05	0.17	0.71	0.31	0.45	0.89	1.54	0.96	1.8 (A100)
VAD † [ 15 ]	Camera	ResNet50	0.54	1.15	1.98	1.22	0.10	0.24	0.96	0.43	0.47	0.83	1.43	0.91	-
SparseDrive †∗ [ 25 ]	Camera	ResNet50	0.44	0.92	1.69	1.01	0.07	0.19	0.71	0.32	0.39	0.77	1.41	0.85	9.0 (RTX4090)
MomAD [ 24 ] †	Camera	ResNet50	0.43	0.88	1.62	0.98	0.06	0.16	0.68	0.30	0.37	0.74	1.30	0.80	7.8 (RTX4090)
UniAD [ 13 ]	Camera	ResNet101	0.45	0.70	1.04	0.73	0.62	0.58	0.63	0.61	0.41	0.68	0.97	0.68	1.8 (A100)

Table 1: Planning results on the nuScenes validation dataset.

\dagger

denotes evaluation protocol used in UniAD.

*

denotes results reproduced with the official checkpoint. As Ref. states, we deactivate the ego status information for a fair comparison.

Method	KBM	ASP	$L_{2}$ (m) $\downarrow$				Col. Rate (%) $\downarrow$				TPC (m) $\downarrow$
Method	KBM	ASP	1s	2s	3s	Avg.	1s	2s	3s	Avg.	1s	2s	3s	Avg.
MomAD (Baseline)	$\cdot$	$\cdot$	0.31	0.57	0.91	0.60	0.010	0.050	0.220	0.090	0.30	0.53	0.78	0.54
Ours (KBM only)	$\checkmark$	$\cdot$	0.27	0.54	0.91	0.58	0.020	0.073	0.238	0.110	0.26	0.46	0.69	0.47
Ours (Neuro-Symbolic)	$\checkmark$	$\checkmark$	0.27	0.54	0.90	0.57	0.010	0.050	0.170	0.075	0.26	0.48	0.66	0.47

Table 2: Component ablation study on the nuScenes validation set. We evaluate the contribution of the Kinematic Bicycle Model (KBM) and ASP-based deductive conditioning.

\checkmark

indicates the module is enabled, while

\cdot

denotes its absence.

Neuro-Symbolic Planning

See all 20 papers

Neuro-Symbolic Drive: Rule-Grounded Faithful Reasoning for Driving VLAs
Jun 22, 2026Xiangbo Gao, Xiukun Huang, Boyu Lu +5VLMs for Autonomous Driving Neuro-Symbolic Reasoning
1Texas A&M University · 2Carnegie Mellon University · University of Maryland +2
KnowDiffuser: A Knowledge-Guided Diffusion Planner with LLM Reasoning
Mar 11, 2026Fan Ding, Xuewen Luo, Fengze Yang +4Autonomous Driving Planning Diffusion-Based Planning
Monash University Malaysia. · University of Utah, USA.
$AutoDrive\text{-}P^3$ : Unified Chain of Perception-Prediction-Planning Thought via Reinforcement Fine-Tuning
Mar 30, 2026Yuqi Ye, Zijian Zhang, Junhong Lin +3Autonomous Driving Planning VLMs for Autonomous Driving
School of Electronic and Computer Engineering, Peking University

Jun 22, 2026J/K move · Enter open · S save

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

Xiangbo Gao, Xiukun Huang, Boyu Lu, Junge Zhang+4

1Texas A&M University · 2Carnegie Mellon University · University of Maryland+2

VLMs for Autonomous Driving Neuro-Symbolic Reasoning Autonomous Driving Planning Neuro-Symbolic Planning

Open paper AskPDF arXiv

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.

Abstract

Neuro-Symbolic Planning

Explore similar work

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

KnowDiffuser: A Knowledge-Guided Diffusion Planner with LLM Reasoning

AutoDrive-P3AutoDrive\text{-}P^3AutoDrive-P3: Unified Chain of Perception-Prediction-Planning Thought via Reinforcement Fine-Tuning

$AutoDrive\text{-}P^3$ : Unified Chain of Perception-Prediction-Planning Thought via Reinforcement Fine-Tuning