Sub-symbolic AI-Enabled RAS

• Sub-symbolic AI-enabled RAS are autonomous systems that leverage deep neural networks, probabilistic models, and hybrid architectures for processing high-dimensional sensor data in real-time.
• They integrate purely sub-symbolic modules with neuro-symbolic hybrids and intersymbolic strategies to balance raw learning with explicit, rule-based reasoning.
• Key challenges include scalability, interpretability, and integration complexity, addressed through formal verification, adaptive uncertainty reduction, and advanced explanation techniques.

Sub-symbolic AI-enabled Robotic and Autonomous Systems (RAS) refer to physical or cyber-physical agents whose core functionalities—such as perception, control, decision-making, and interaction—are driven by AI techniques operating in a distributed, continuous, and high-dimensional manner, rather than via explicit rule-based symbolic computation. The haLLMark of sub-symbolic AI in RAS is the deployment of machine learning (primarily deep neural networks), probabilistic learning, and hybrid neuro-symbolic architectures at the level of raw sensor signals, with the dual goals of robust real-time operation in dynamic environments and, increasingly, context-driven explainability and trustworthiness.

1. Core Principles and Motivations

Sub-symbolic AI methods, encompassing deep neural networks, support vector machines, and probabilistic models, are embedded into RAS to process high-dimensional sensory inputs (e.g., camera, LiDAR, radar, audio). Unlike symbolic AI, where knowledge is encoded in interpretable, rule-based formalisms (e.g., logic, ontologies), sub-symbolic systems learn statistical regularities from data, resulting in distributed representations where individual model components (weights, neurons) lack standalone interpretability (He et al., 2021).

The principal motivation for employing sub-symbolic AI in RAS lies in the ability to handle the scale and complexity of real-world sensory data, perform robust pattern recognition and perception under noise and uncertainty, and enable adaptive learning without manual feature engineering. Sub-symbolic models can generalize from experience, extract latent structure from unstructured data, and react rapidly to novel scenarios essential in domains such as autonomous driving, service robotics, and industrial automation (He et al., 2021, Alt et al., 21 Apr 2024).

2. System Architectures and Integration Strategies

Sub-symbolic AI in RAS manifests primarily in three integration paradigms:

• Purely Sub-symbolic Modules: Perception, control, and decision-making units based solely on neural or probabilistic models (e.g., deep convolutional networks for vision, neural path planning for navigation) (He et al., 2021).
• Neuro-Symbolic Hybrids: Architectures that combine sub-symbolic learning with symbolic knowledge representation, such as integrating neural perception with knowledge graph embeddings (TransE, HolE) to facilitate context understanding; architectures may include explicit modules for symbolic knowledge injection via attention mechanisms or knowledge graph embeddings (e.g., $h + r \approx t$ relation in TransE embeddings) (Oltramari et al., 2020, Sheth et al., 2023).
• Intersymbolic and Cognitive Systems: Advanced integration schemes interlink symbolic logic (for safety barriers, mission constraints, or formal verification loops) with sub-symbolic controllers, as in ModelPlex shielding or formal synthesis techniques (Platzer, 17 Jun 2024, Azaiez et al., 24 Sep 2025). Cognitive neuro-symbolic systems may connect classical architectures (e.g., ACT-R) with neural modules, creating bidirectional links between explicit knowledge and latent learned representations (Oltramari, 2023).

Integration Paradigm	Sub-symbolic Role	Symbolic Role
Pure Sub-symbolic	Perception, control, planning	None
Neuro-symbolic Hybrid	Latent representation, adaptation	Contextual grounding, rules
Intersymbolic/Cognitive	Sensorimotor learning, policy learning	Verification, explicit reasoning

Hybrid strategies are central to addressing the interpretability and safety limitations prevalent in sub-symbolic approaches (Oltramari et al., 2020, Wan et al., 2 Jan 2024).

3. Application Domains and Representative Workflows

Sub-symbolic AI-enabled RAS operate across varied domains with workflows generally comprising the following elements:

• Sensor Fusion and Perception: Multi-modal sensor data are fused and processed via neural or probabilistic models. For example, object detectors (e.g., Faster R-CNN, YOLO) parse camera/LiDAR inputs to produce bounding boxes and semantic labels; joint perception methods combine these via scene graph generation (SGG) for rich context modeling (Hallyburton et al., 27 May 2025, Oltramari et al., 2020).
• Context Understanding and Situation Assessment: Neural models map perceived features into latent spaces; in neuro-symbolic systems, these are further matched or “grounded” against knowledge graphs through embeddings, enabling context inference (e.g., scene matching via cosine similarity in KGE space) (Oltramari et al., 2020, Piplai et al., 2023).
• Decision Making and Control: Sub-symbolic policies or controllers compute actions, optionally constrained by symbolic rules (e.g., safety envelopes from a knowledge base or rules injected via regularization terms: $$L_\text{total} = L_\text{task} + \lambda L_\text{logic}$$) (Colelough et al., 9 Jan 2025, Ciatto et al., 23 Jan 2025).
• Explanation and Trust Management: To address the black-box nature, explainability mechanisms extract symbolic rules from neural models (symbolic knowledge extraction, SKE), or inject domain knowledge to modify learned policies (symbolic knowledge injection, SKI). Techniques include GNNExpainer-DL-Learner pipelines, rule extraction meta-models, and fidelity metrics to verify correspondence between explanations and model decisions (Himmelhuber et al., 2021, Himmelhuber et al., 2022, Ciatto et al., 23 Jan 2025).
• Formal Verification and Synthesis: For safety-critical applications, formal methods (e.g., probabilistic model checking, synthesis of uncertainty reduction controllers) are increasingly integrated to assure that sub-symbolic modules operate within provable safety and performance bounds (Azaiez et al., 24 Sep 2025).

4. Interpretability, Explainability, and Trustworthiness

Addressing the opacity of sub-symbolic AI is a prominent research focus:

• Symbolic Knowledge Extraction (SKE): Decompositional or pedagogical methods distill human-understandable rule sets, logic expressions, or decision trees from neural predictors, enabling post-hoc auditing and debugging. Taxonomies classify SKE by access level (white-box/black-box), input data modality, and output expressiveness (Ciatto et al., 23 Jan 2025, Himmelhuber et al., 2021).
• Symbolic Knowledge Injection (SKI): Domain knowledge (safety constraints, commonsense rules, regulatory constraints) is incorporated into sub-symbolic models via predictor structuring, knowledge embedding, or regularization of loss functions (e.g., $$L_\text{total} = L_\text{data} + \lambda L_\text{constraint}$$) (Ciatto et al., 23 Jan 2025, Piplai et al., 2023).
• Fidelity Metrics: Quantitative measures compare the coverage of symbolic explanations against sub-symbolic decision rationales (e.g., the overlap between GNNExplainer subgraphs and symbolic explainer classes), with high fidelity denoting faithful and trustworthy model introspection (Himmelhuber et al., 2021, Himmelhuber et al., 2022).
• Responsible AI and Human Acceptance: Trust in RAS is now recognized as dependent not only on technical performance (worthiness) but also on the demonstrable reliability, safety, and explainability (trustiness), with user acceptance contingent on both pillars (He et al., 2021, Hooshyar et al., 1 Apr 2025).

5. Systemic Challenges and Engineering Trade-offs

Development and deployment of sub-symbolic AI-enabled RAS involve numerous challenges:

• Scalability: Embedding large-scale knowledge representations (e.g., NuScenes KG with >2 million entities) and maintaining real-time performance in resource-constrained robotic platforms are significant hurdles (Oltramari et al., 2020, Wan et al., 2 Jan 2024).
• Integration Complexity: Cross-modal fusion of symbolic (e.g., traffic laws, mission objectives) and sub-symbolic (sensor feature spaces) representations requires robust latent space alignment and attention mechanisms.
• Data and Domain Alignment: Injected symbolic knowledge must be domain-relevant and appropriately matched to the data; misaligned knowledge bases (e.g., pre-training on unrelated procedural knowledge) can degrade performance (Oltramari et al., 2020, Piplai et al., 2023).
• Interpretability vs. Performance: Increased transparency may introduce computational overheads or restrict model complexity, potentially affecting state-of-the-art accuracy (Oltramari et al., 2020, Colelough et al., 9 Jan 2025).
• Verification and Safety Assurance: Formal methods, including probabilistic model checking and formal synthesis of controllers, support “correct-by-construction” guarantees, but may increase design complexity and require sophisticated tool support (Azaiez et al., 24 Sep 2025, Platzer, 17 Jun 2024).

6. Emerging Trends and Future Directions

Forward-looking research targets several areas to further advance sub-symbolic AI in RAS:

• Standardization and Taxonomies: Systematic reviews now classify methods for symbolic knowledge extraction/injection, neuro-symbolic integration, and cognitive architecture coupling, facilitating method selection and engineering best practices (Ciatto et al., 23 Jan 2025, Colelough et al., 9 Jan 2025, Wan et al., 2 Jan 2024).
• Meta-cognition and Cognitive Integration: Incorporating supervisory layers that enable systems to monitor and adjust their own reasoning (meta-cognition) is identified as a key driver for adaptivity and trust in autonomous systems (Colelough et al., 9 Jan 2025, Oltramari, 2023).
• Hybrid Hardware/Software Platforms: Emerging hardware-software stacks are tailored for heterogeneous workloads, supporting both dense neural computations and symbolic logic/rule reasoning (Wan et al., 2 Jan 2024).
• Adaptive Uncertainty Reduction: Integration of uncertainty reduction controllers (synthesized via probabilistic formal methods) into RAS architectures establishes adaptive, information-driven operation that proactively addresses environmental uncertainty (Azaiez et al., 24 Sep 2025).
• Scalable Explainability: Continued development of high-fidelity explanation techniques (fidelity metrics, human-centric causal reasoning output) for varied AI architectures ensures transparency remains a priority even as models increase in complexity (Himmelhuber et al., 2022, Himmelhuber et al., 2021).
• Integration of LLMs and In-context Reasoning: Current paradigms employ LLMs and vector-symbolic architectures for scalable, implicit symbolic reasoning, enabling rapid task adaptation and flexible decision-making (Xiong et al., 11 Jul 2024, Griffiths et al., 7 Aug 2025).

7. Summary Table: Illustrative Approaches in Sub-symbolic AI-enabled RAS

Challenge/Goal	Sub-symbolic Technique	Hybrid/Symbolic Remedy
Scene Recognition	DNNs, GNNs, sensor fusion	Knowledge graphs, KG embedding, attention
Safe Control	Reinforcement learning, learned policies	ModelPlex, logical rules, formal synthesis
Explainability	SHAP, feature attribution	SKE (rule extraction), fidelity metrics
Cybersecurity/Anomaly	ML on GNNs (anomaly detection)	Symbolic XAI (DL-Learner, ontologies)
User Trust/Acceptance	Performance-driven, opaque models	Dual symbolic-sub-symbolic architectures

This multidimensional synthesis reflects the current state of the field: Sub-symbolic AI methods are fundamental to RAS due to their adaptability and representational power, but their limitations have led to a new generation of hybrid neuro-symbolic and intersymbolic systems that combine data-driven learning with explicit, explainable, and safety-assured symbolic reasoning (Oltramari et al., 2020, He et al., 2021, Colelough et al., 9 Jan 2025, Platzer, 17 Jun 2024, Hallyburton et al., 27 May 2025).