Real-Time Reachability for Neurosymbolic Reinforcement Learning-based Safe Autonomous Navigation

Nicholas Potteiger, Diego Manzanas Lopez, Taylor T. Johnson, Xenofon Koutsoukos
Proceedings of the International Conference on Neuro-symbolic Systems, PMLR 288:104-126, 2025.

Abstract

Safety is essential in autonomous navigation, especially as autonomous systems deploy to new environments where collision avoidance is critical. Neurosymbolic reinforcement learning (NeSy RL) approaches show promise for advancing long-term navigation by using symbolic planners to compute high-level waypoints and goal-conditioned RL for low-level control. However, ensuring safety within these frameworks remains a challenge, particularly in new environments that the agent was not optimized for. Current safe RL-based navigation techniques offer robust frameworks for ensuring safety. However, these approaches are not adapted for NeSy RL and also present challenges: they can be computationally intensive or constrained by conservative control. To overcome these limitations, we propose a novel approach to safely and efficiently navigate a NeSy RL agent in new environments. The proposed method uses real-time reachability analysis to select subgoals between waypoints and safeguard the actions of a goal-conditioned RL policy. We implement the approach in Rust and develop a software package, RusTReach, for real-time reachability analysis. We deploy our approach on an embedded device and compare against four approaches in a long-term quadcopter navigation task in a new environment. Our evaluation reveals that our approach is at least 1.7 times faster at navigating than a state-of-the-art alternative while maintaining safety and real-time constraint compliance. Code and videos available at https://github.com/npotteig/rustreach.

Cite this Paper

BibTeX
@InProceedings{pmlr-v288-potteiger25a, title = {Real-Time Reachability for Neurosymbolic Reinforcement Learning-based Safe Autonomous Navigation}, author = {Potteiger, Nicholas and Manzanas Lopez, Diego and Johnson, Taylor T. and Koutsoukos, Xenofon}, booktitle = {Proceedings of the International Conference on Neuro-symbolic Systems}, pages = {104--126}, year = {2025}, editor = {Pappas, George and Ravikumar, Pradeep and Seshia, Sanjit A.}, volume = {288}, series = {Proceedings of Machine Learning Research}, month = {28--30 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v288/main/assets/potteiger25a/potteiger25a.pdf}, url = {https://proceedings.mlr.press/v288/potteiger25a.html}, abstract = {Safety is essential in autonomous navigation, especially as autonomous systems deploy to new environments where collision avoidance is critical. Neurosymbolic reinforcement learning (NeSy RL) approaches show promise for advancing long-term navigation by using symbolic planners to compute high-level waypoints and goal-conditioned RL for low-level control. However, ensuring safety within these frameworks remains a challenge, particularly in new environments that the agent was not optimized for. Current safe RL-based navigation techniques offer robust frameworks for ensuring safety. However, these approaches are not adapted for NeSy RL and also present challenges: they can be computationally intensive or constrained by conservative control. To overcome these limitations, we propose a novel approach to safely and efficiently navigate a NeSy RL agent in new environments. The proposed method uses real-time reachability analysis to select subgoals between waypoints and safeguard the actions of a goal-conditioned RL policy. We implement the approach in Rust and develop a software package, RusTReach, for real-time reachability analysis. We deploy our approach on an embedded device and compare against four approaches in a long-term quadcopter navigation task in a new environment. Our evaluation reveals that our approach is at least 1.7 times faster at navigating than a state-of-the-art alternative while maintaining safety and real-time constraint compliance. Code and videos available at https://github.com/npotteig/rustreach.} }
Endnote
%0 Conference Paper %T Real-Time Reachability for Neurosymbolic Reinforcement Learning-based Safe Autonomous Navigation %A Nicholas Potteiger %A Diego Manzanas Lopez %A Taylor T. Johnson %A Xenofon Koutsoukos %B Proceedings of the International Conference on Neuro-symbolic Systems %C Proceedings of Machine Learning Research %D 2025 %E George Pappas %E Pradeep Ravikumar %E Sanjit A. Seshia %F pmlr-v288-potteiger25a %I PMLR %P 104--126 %U https://proceedings.mlr.press/v288/potteiger25a.html %V 288 %X Safety is essential in autonomous navigation, especially as autonomous systems deploy to new environments where collision avoidance is critical. Neurosymbolic reinforcement learning (NeSy RL) approaches show promise for advancing long-term navigation by using symbolic planners to compute high-level waypoints and goal-conditioned RL for low-level control. However, ensuring safety within these frameworks remains a challenge, particularly in new environments that the agent was not optimized for. Current safe RL-based navigation techniques offer robust frameworks for ensuring safety. However, these approaches are not adapted for NeSy RL and also present challenges: they can be computationally intensive or constrained by conservative control. To overcome these limitations, we propose a novel approach to safely and efficiently navigate a NeSy RL agent in new environments. The proposed method uses real-time reachability analysis to select subgoals between waypoints and safeguard the actions of a goal-conditioned RL policy. We implement the approach in Rust and develop a software package, RusTReach, for real-time reachability analysis. We deploy our approach on an embedded device and compare against four approaches in a long-term quadcopter navigation task in a new environment. Our evaluation reveals that our approach is at least 1.7 times faster at navigating than a state-of-the-art alternative while maintaining safety and real-time constraint compliance. Code and videos available at https://github.com/npotteig/rustreach.
APA
Potteiger, N., Manzanas Lopez, D., Johnson, T.T. & Koutsoukos, X.. (2025). Real-Time Reachability for Neurosymbolic Reinforcement Learning-based Safe Autonomous Navigation. Proceedings of the International Conference on Neuro-symbolic Systems, in Proceedings of Machine Learning Research 288:104-126 Available from https://proceedings.mlr.press/v288/potteiger25a.html.

Related Material