In this paper, we analyze the role of Neuro-Symbolic (NS) methods within the standard Observe, Orient, Decide and Act (OODA) framework and discuss their associated opportunities and challenges. To ground the discussion, we consider the OODA loop applied to a wildland firefighting use case where mixed teams of autonomous agents and humans need to collaboratively work to contain and ultimately extinguish large-scale fires. NS methods are appealing as they enable the integration of symbolic knowledge coming, for example, from time-tested firefighting tactics and procedures encoded in training manuals. However, there are key challenges in capturing and integrating such information within NS pipelines, especially in a way that is scalable with respect to the number of procedures, dynamics of the environment, and regulations. The integration of such symbolic priors within a NS OODA loop will enable autonomous systems to generate highly expressive yet compressed representations (or abstractions) of the environment and other agents' state, actions, and intent. However, NS integration brings new challenges, such as, designing abstractions that are composable and transferable across complex tasks, deciding how to ground symbols to sensory data, and verifying their validity to ensure they lead to safe decisions. Symbolic abstractions also need to deal with the fact that sensory data is uncertain-like the behavior of the fire front (boundary, rate of spread, etc.). Furthermore, prior knowledge may be incomplete and inconsistent leading to representations that need to be capable of explicitly modelling and managing ambiguity. Addressing such challenge problems will not only advance the state-of-the-art in NS AI, but also concretely demonstrate the benefits of such methods within the autonomy domain.
In recent years, deep neural networks (DNNs) and embedded GPUs have significantly improved the
capabilities of an autonomous system across the full Observe, Orient, Decide and Act (OODA) loop [
Consider, for example, the wildland firefighting problem shown in Figure 1. In this type of complex mission, there are many challenges in developing an understanding of the current and future state of the fire supporting the generation of strategies and actions to effectively fight a wildland fire.
As the figure shows, fires may be in various states, burning at different rates along challenging terrains; winds and vegetation fuel the fires in a difficult to predict way and thick clouds of smoke make fire front detection and overall flight arduous. Furthermore, risks to people and infrastructure make decisions on how to tackle the fires difficult.
Purely data-driven machine learning (ML) approaches may not be sufficient to tackle the challenges of such missions. For example, we expect that deployed autonomous systems will: coordinate with humans and follow specific set of tactics and guidelines, build and reason over large-scale spatio-temporal models of the world, adapt online to new situations for which training data was not available before the mission, and manage both uncertainty and ambiguity.
Neuro-Symbolic (NS) methods provide a new framework to overcome some of the limitations of
purely data-driven ML approaches [
The challenges discussed in the following sections stem from the authors’ recent experience in developing and integrating NS methods within autonomous systems. Not all of them are exclusive to NSenabled autonomy. 2.1.
NS methods enable the integration of symbolic priors during both the learning and inference. In wildland firefighting, for example, symbolic priors can be a set of well-established tactics and procedures when fighting fires (e.g., never attack an aggressive fire at the head or up-slope from an active fire front) as well as priorities on how tactics should be selected (e.g., preservation of life being at the highest priority). Symbolic knowledge also imposes requirements on the “Observe” and “Orient” parts of the loop as certain types of specific features in the terrain, vegetation or land usage need to be detected and characterized.
Within the Decide portion of the loop, symbolic priors impact decisions as the autonomous systems
need to conform to procedures to both coordinate effectively with humans and for ethical and legal
reasons. ML methods, such as DNNs, require a lot of data to learn to recognize such complex behavioral
patterns and from there select the appropriate actions. NS methods such as [
However, at this stage there does not seem to exist well-established quantitative metrics that help developers understand the tradeoff between data requirements, richness of symbolic priors and performance of a NS system. In the application domain this is ultimately translating into development costs.
For example, in the wildland firefighting setting, generating (or collecting) more data, and possibly labeling the data, can be costly. Transforming firefighting procedures into ontologies and knowledge graphs by ingesting, for example, procedure manuals that encode information in highly unstructured forms, such as pictures and/or sketches can also very onerous.
One of the key enablers of the success of supervised learning has been the availability of extremely
large sets of labelled data, typically released as benchmark data. Many existing NS methods (for
example, [
The development of advanced Large Language Models (LLMs) [
In the context of symbolic knowledge capture, there is a more fundamental question that we believe
should be pursued more aggressively. Generally, formal knowledge is encoded via description logics,
and this raises the question if this formalism is sufficient or even optimal for NS methods. Recent
developments [
Symbols represent concepts ranging from objects and classes to tasks and plans or from beliefs to properties and strategies. To apply symbolic knowledge, a NS system must be able to ground a symbol, i.e., associate it (possibly probabilistically) with real world observations or with combinations of other related symbols. Often symbols are arranged hierarchically (e.g., in an ontology) to represent concepts at many levels of abstraction, so we refer to the association of symbols with real world observations or with combinations of other related symbols as an abstraction. Hierarchical abstractions occur naturally in the OODA loop, where “Observe” creates symbols representing objects in the environment, “Orient” groups these observed objects and relates them to each other and to higher level concepts through a rich set of relationships, “Decide” plans actions based on the hierarchical plans and goals of the agents and symbolic representations, and “Act” carries out the actions in context of the observed objects and agents and the expected plans and goals of the agents. Prior knowledge generally defines abstractions, but common sense and efficient reasoning requires an ability to fluidly relate existing abstractions to each other, to refine them based on context, and to create new abstractions for new situations. Abstractions should have, at a minimum, the following characteristics: 1) Utility: the abstraction should be compositional, modular, and reusable/transferable (ideally across applications and domains). Reusability may require adaptation of the abstraction to new domains/contexts. 2) Interpretability: the abstractions should be defined in terms of existing abstractions/symbols and observations in a way that humans can understand and such that it is easy to determine whether a given abstraction is applicable in a given context.
In the firefighting domain, abstractions within the “Orient” part of the OODA loop will be designed to support the efficient prediction of wildfire behavior from the fire’s current state, the fuel characteristics, the topography, and the weather, and to characterize the risk to structures and human life in planning the firefighting effort.
Within “Decide”, the abstractions will likely need to align with the hierarchical organization of the command structure. At the top is the Incident Commander (IC) who makes both tactical and strategic decisions about how to best fight the fire that has both a long-term and short-term effect. At the lower layers, the abstractions will capture tactics and actions other agents, both autonomous and humans, are going to take to carry out the IC decisions.
Although human experts, field manuals and fire behavior models [
A key challenge to take advantage of NS methods in developing next generation OODA loops is on developing theoretical and computational methods for generating and adapting abstractions across the elements of the loop as well as measuring their utility and interpretability. Each of these characteristics is critical to ensure consistency of the symbolic representations across abstractions, the management of uncertainty and ambiguities (see Section 2.3) and to support reasoning. We believe that such abstractions should be designed to ensure transferability across problem instances/types and scalable to large number of heterogenous systems. 2.3.
Although integration of probabilistic models, to consider, e.g., sources of uncertainty, have been
considered in the context of NS methods [
Although, we believe it is appealing to consider information theoretic approaches to study
quantization tradeoffs, in the context of autonomous systems using NS components, the choice of symbols is
not just a “bottom-up” process, from continuous to discrete domains, but also a “top-down” process as
the symbols have associated semantics that depend on the mission, external knowledge encoded in
ontologies and, potentially, the need to be interpretable by humans. In the firefighting setting, if we
require autonomous systems to coordinate with humans, it seems highly desirable to have symbolic
representations that match humans’ representations. Although there is initial work in developing
interpretable symbolic NS world models for autonomous systems [
A different, but related, important aspect in the design of a NS OODA loop is the need to ensure that symbol groundings do not conflict across the components of an OODA loop. The use of different complex sensors (e.g., EO/IR cameras, LIDARs, etc.) within the “Observe” and the generation of multiple abstractions in “Orient” to tradeoff computation complexity and accuracy, combined with the fact that the world is only partially observable, can lead to situations where symbolic representations are inconsistent. We believe that NS OODA loops will require new approaches to represent and reason under such uncertainty and ambiguity without sacrificing reliability and assurability, see also Section 2.4. 2.4.
The deployment of OODA loop solutions to safety critical domains, such as wildland firefighting require demonstrably correct, reliable, and safe behavior in all foreseeable situations. In regulated industries users need explainability, transparency, and auditability of the system’s behavior [24]. We call this the assurance challenge. Assured systems aim to satisfy four Overarching Properties (OAs) listed below [25]: • Intent: The system's defined intended behavior must be correct and complete with respect • • • to the desired behavior.
Correctness: The implementation must be correct with respect to its defined intended behavior under foreseeable operating conditions.
Innocuity: Any part of the implementation that is not required by the defined intended behavior must not have unacceptable impact.
Operation: The system must possess mechanisms for addressing correctness or intent deficiencies and for mitigating unacceptable impacts manifested during operation.
The verification of a system is meant to establish whether OODA components satisfy the four OAs
via objective evidence, thus providing confidence that a system has been “built right”. The symbolic
interfaces and knowledge representations afforded by NS techniques, compared to DNN based
approaches, enable that possibility. For instance, Logic Tensor Networks (LTN) [
Although the assurance of NS OODA loops can take advantage of explicit symbolic representations these may be still learned from combination of data and explicit knowledge.
Learned NS functions implies the presence of uncertainty and ambiguity, as mentioned before, that not only needs to be managed but also treated as an assurance gap. Seeding learning by injecting symbolic priors may accelerate learning but also introduces grounding boundary discrepancies, in other words, a discordance between real world objects and their representations. Furthermore, the validity of learned symbols and their symbolic relations are highly dependent on the quality, quantity, and relevance of the input and training data, and thus connected to the challenges in Section 2.1 and 2.2. More generally, we believe that the connections and interfaces between the symbolic and the sub-symbolic components of NS systems need to be better understood from an assurance perspective and, likely, new mitigation approaches need to be developed to provide strong safety guarantees.
NS methods can greatly improve the way we design and deploy autonomous systems in complex missions. The opportunity to incorporate prior symbolic knowledge, integrate deep learning with symbolic representations and reasoning, provide interpretability and enable formal analysis are huge advantages compared to approaches that completely rely on statistical machine learning. There are, however, challenges that we need to overcome to design next generation of NS based autonomous systems. In this paper we have listed a few challenges related to the tradeoff between data requirements and symbolic knowledge in training NS architectures, the development of composable symbolic abstraction for reasoning and the need of formalize methods for managing uncertainty and ambiguity in these novel architectures. We believe that autonomous systems/robotics can be a great application domain to drive the development of new theoretical and computational advances in NS learning and reasoning. Specifically, we have also introduced a rich problem domain, wildland firefighting, that provides the research community with a strong and concrete motivating example where we believe NS methods can make an important contribution and impact.
[23] A. Rosinol, A. Gupta, M. Abate, J. Shi, and L. Carlone, “3D Dynamic Scene Graphs: Actionable Spatial Perception with Places, Objects, and Humans”. In Robotics Science and Systems, 2020. [24] Z. Iqbal, “Assurance of Machine Learning/TinyML in Safety-Critical Domains”. IEEE Symposium on Visual Languages and Human-Centric Computing (VL/HCC), 2022. [25] C. M. Holloway, “Understanding the Overarching Properties”. National Aeronautics and Space Administration, Langley Research Center, 2019. Online: http://shemesh.larc.nasa.gov/ arg/utops.pdf. [26] R. Socher, D. Chen, C. D. Manning, and A. Ng, “Reasoning with Neural Tensor Networks for Knowledge Base Completion”. In Advances in Neural Information Processing Systems (NeurIPS), 2013. [27] P. Smolensky, M. Lee, X. He, W.-t. Yih, J. Gao, and L. Deng, “Basic Reasoning with Tensor
Product Representations”. ArXiv, https://arxiv.org/abs/1601.02745, 2016. [28] Q. Cao, W. Wan, X. Liang, and L. Lin, “Graph Reasoning Networks and Applications,” in Neuro-Symbolic Artificial Intelligence: The State of the Art, P. Hitzler and M. K. Sarker, Eds. IOS Press, 2022. [29] Z.-H. Zhou and Y.-X. Huang, “Abductive Learning,” in Neuro-Symbolic Artificial Intelligence: The State of the Art, P. Hitzler and M. K. Sarker, Eds. IOS Press, 2022.