Toward Assured Autonomy Using Neurosymbolic Components and Systems

Abhishek Dubey, Taylor T. Johnson, Xenofon Koutsoukos, Baiting Luo, Diego Manzanas Lopez, Miklos Maroti, Ayan Mukhopadhyay, Nicholas Potteiger, Serena Serbinowska, Daniel Stojcsics, Yunuo Zhang, Gabor Karsai

Neurosymbolic AI 2026

PDF DOI

Why This Matters

Autonomous systems increasingly rely on neural components for perception and decision-making, but assuring the safety of these components remains a fundamental challenge. Pure neural approaches lack formal guarantees, while pure symbolic approaches cannot handle the complexity of real-world perception. The innovation is decomposing the autonomy stack into neurosymbolic components where each combines learned perception or prediction with symbolic reasoning and constraints, and then applying model checking to verify properties of the resulting hybrid system — providing a principled path toward assured autonomy.

What We Did

This book chapter presents how neurosymbolic techniques can implement three core functions of an autonomous UAV system: world model maintenance (updating an internal representation of the environment from sensory inputs), planning (generating waypoints for the vehicle), and trajectory control (producing fine-grain control commands). The components are developed for a UAV mission — localizing a specific object in an urban area — and evaluated in a virtual environment. An assurance technique based on model checking is also presented for verifying neurosymbolic components that combine finite-state control with neural modules.

Key Results

The neurosymbolic components successfully implement world model maintenance, subgoal-based planning, and trajectory control for a UAV target localization mission. Evaluation in a virtual urban environment demonstrates that the neurosymbolic architecture achieves mission objectives while enabling formal verification of safety properties through model checking. The chapter documents lessons learned from integrating neural and symbolic components, including the importance of safety constraints in the planning loop and the role of landmark selection in maintaining accurate world models.

Full Abstract

Cite This Paper

@inbook{dubey2026neurosymbolic,
  author = {Dubey, Abhishek and Johnson, Taylor T. and Koutsoukos, Xenofon and Luo, Baiting and Lopez, Diego Manzanas and Maroti, Miklos and Mukhopadhyay, Ayan and Potteiger, Nicholas and Serbinowska, Serena and Stojcsics, Daniel and Zhang, Yunuo and Karsai, Gabor},
  title = {Toward Assured Autonomy Using Neurosymbolic Components and Systems},
  booktitle = {Neurosymbolic AI},
  publisher = {John Wiley \& Sons, Ltd},
  year = {2026},
  chapter = {4},
  pages = {89-118},
  isbn = {9781394302406},
  doi = {https://doi.org/10.1002/9781394302406.ch04},
  url = {https://onlinelibrary.wiley.com/doi/abs/10.1002/9781394302406.ch04},
  abstract = {Neurosymbolic techniques are expected to deliver more functionalities and better performance in autonomous systems, but their assurance remains a challenge. There are various roles such components can play in an autonomous vehicle, for instance, world model maintenance, planning, and trajectory control. The world model is an internal representation of the external environment of the vehicle that is continuously updated based on new sensory inputs; the planning component generates waypoints for the vehicle to reach, while the trajectory controller produces the fine-grain control commands for the vehicle's path. This chapter presents how these three functions can be implemented using neurosymbolic techniques, and presents results and the lessons learned. The components were developed in the context of a UAV executing a specific mission: localization of a specific object in an urban area, and evaluated in a virtual environment. An assurance technique based on model checking is presented that can be applied to a class of neurosymbolic components that include finite-state control with neural components.},
  keywords = {neurosymbolic AI, assured autonomy, UAV, world model, planning, trajectory control, model checking, hybrid systems}
}
Quick Info
Year 2026
DOI https://doi.org/10.1002/978...
Keywords
neurosymbolic AI assured autonomy UAV world model planning trajectory control model checking hybrid systems
Research Areas
planning CPS scalable AI
Search Tags

Toward, Assured, Autonomy, Neurosymbolic, Components, Systems, neurosymbolic AI, assured autonomy, UAV, world model, planning, trajectory control, model checking, hybrid systems, CPS, scalable AI, 2026, Dubey, Johnson, Koutsoukos, Luo, Lopez, Maroti, Mukhopadhyay, Potteiger, Serbinowska, Stojcsics, Zhang, Karsai