Minimum Viable Autonomy

Considerations for the application of autonomy in robotic and autonomous systems

Introducing Minimum Viable Autonomy for robotic systems

The idea of Minimum Viable Autonomy (MVA) is an emerging concept in the development and deployment of autonomous systems, contextualised for intended system applications to realise a desired outcome. I'll start by providing an intuitive definition for MVA, given as

The minimum amount of autonomous functionality required for a system to perform a task or set of tasks. 

This article will aim to coherently define MVA, explore its relevance, and establish the necessary chain of logic for its use. I'll include various academic and applied sources focusing on robotics and autonomous systems to position the concept. 

A short history

A quick search on Scopus or Google Scholar for MVA returns few to no similar uses of the term 'minimum viable autonomy'. A broader search highlights that the exact term may have first appeared in a 2020 edition of Aviation Week magazine, noting in a single use that 'minimum viable autonomy is breaking the problem down into bite-sized chunks that you can push through the certification process' [1]. 

When prompted with 'what is minimum viable autonomy', ChatGPT-4 suggests that MVA is a concept in the development of autonomous systems, drawing from the idea of the MVP in product development,' going on to state that 'in essence, MVA is about finding the right balance between too little autonomy, where the system cannot operate independently, and too much, where the system may become unnecessarily complex or difficult to manage.' This is a pretty good summary and captures the essence of the idea. 

Relaxing the search criteria to consider 'minimum autonomy' yields far more significant results. However, it diverges from our application context with results ranging from the medical field to international relations. Relevant literature of this broader term appeared in-earnest in the early 2000s, with research published on the teleoperation of humanoid robots [2] and frameworks to control interactive and tele-operative systems via the internet [3]. Defining minimum autonomy remains opaque, with few studies clearly defining the term. 

In the studies mentioned above and more from other researchers, the use of minimum autonomy is varied. This spans from representing the physical characteristics of an uncrewed system, such as the required operating time or distance to be traversed, to a catch-all term for cognitive function aspects of a platform. 

While aspects of the intuitive definition are present as subordinate themes within the literature, a concrete definition to share understanding and move forward collaboratively is yet to be presented.

Enabling concepts and framing 

Before further motivating the need for MVA and the logic for a revised definition, I'll discuss what I consider to be the tenets of MVA and highlight the position of this idea and the influences that enable it. I'm positioning MVA at the intersection of several concepts across a spectrum of applied and theoretical fields. It is essential to highlight the connective function of MVA relevant to this work and clearly outline where we're going from this foundation.

Minimum Viable Autonomy relies on five enabling tenets that position the concept.

Minimum Viable Product (MVP)

Emerging in the early 2000s for product development, the MVP broadly describes the simplest product form that can be released to meet minimum requirements. The MVP is important as a conceptual aiming mark to consider the developmental and implementation pathways of autonomous systems and how these interoperate with humans. 

Following from the earlier response of ChatGPT-4, MVA is a linguistic derivation of MVP. The concepts share a common philosophical origin, advocating for the simplest, most fundamental, and most effective version. Where MVA does depart conceptually from MVP is the outcome focus. MVA is capability-focused, whereas the MVP is traditionally customer-focused.

Minimum Autonomy

Minimum autonomy in our context comprises the atomic components of an autonomous system that facilitate the conduct of sensing, deciding and acting processes without regard to performance or efficiency. Minimum autonomy here focuses on the system's functional operation, instead of the intended system use. 

MVA specifies a minimum autonomy in the context of an operational setting to accomplish tasks effectively and reliably. While minimum autonomy provides the building blocks for an autonomous system, MVA ensures the autonomous functions are sufficiently mature to complete a mission. 

Context

Researchers have defined context for autonomous vehicles as the minimum information required to operate autonomously and achieve the mission [4]. Two aspects of this definition are important for MVA. The first is minimum information, which has inherent linkages to Shannon's Information Theory [5], and the second is the connection to an autonomous system's mission outcome (objective).

Framing MVA through this context lens helps bridge an autonomous system's input space (information received) and the desired outcome (mission success). I've discussed earlier that the concept of MVA is deeply intertwined with the intended operational use. Context is the dominant feature that modulates an autonomous system's viability for a given use case. 

Predictability

Predictability broadly refers to anticipating or estimating a system's actions or outputs, introduced as a tenet of transparency in human-machine teaming [6]. The key idea here is that a human must be able to trust that a system will perform within desired characteristic bounds, such that the system's behaviours are reliable, verifiable and repeatable within a mission. The International Organization for Standardization ISO 31000 Standard [7] defines risk as 'the effect of uncertainty on objectives.' Deviation represents increased risk between how a human would conduct a mission and how an autonomous system conducts it. 

Predictability as a tenet inherently couples how well a human can anticipate or estimate the outputs of an autonomous system. This connection is essential to develop bi-directional trust that promotes effective and efficient collaboration between humans and autonomous systems. Through this lens, MVA ensures that an autonomous system is not overly complex for a given context. 

Stability

Originating from strategic analysis, particularly about nuclear strategy and proliferation [8], stability is the final enabling concept of MVA. Late in 2023, the UK House of Lords report titled Proceed with Caution: Artificial Intelligence in Weapon Systems [9] discussed a rationale for stability in the context of autonomous systems. For MVA, stability refers to the idea that employing an autonomous system should not lead to a disruptive (destabilising) outcome for the task. 

Minimising the risk of unintended outcomes (destabilising) connects the tenet of predictability for the operational use context. This is notably important when the consequences of autonomous system actions are substantial, such as in military, security, and emergency service applications. 

Further discussing high-risk settings within the military, mission command is an overarching philosophy for conducting military actions [10], espousing an outcomes-based approach without specifying how particular actions are achieved. Through the stability principle lens, MVA allows for an appropriate level of autonomy that enables an autonomous system to act decisively and effectively while ensuring actions are reliable for the operational use context. You could argue that MVA provides a technological dimension to the mission command philosophy, enhancing capability deliberately and considerately.

The need for minimum viable autonomy

The central premise of MVA is based on the idea that the capabilities of an autonomous system should be commensurate to the mission objectives and operational context that the system must achieve. Supporting this is the general trend that higher autonomy leads to a more extensive decision space as the autonomous system has more possible behaviours to execute. Researchers have highlighted that as system autonomy increases, this can lead to heightened technostress for humans [11], diminishing potential benefits across various scenarios. 

As an outcome, the implication is that increased technostress could lead to a lack of human control or oversight, increased perceived workloads to counteract possible adverse autonomy outcomes, and decreased mission task effectiveness. When employed within contexts possessing negative outcome possibilities (high potential risk), little is known about how the interactions between humans and autonomous systems manifest to influence a human's decision-making ability [12]. 

The good news is that solutions presently exist to help counteract potentially harmful outcomes, with a range of options presented that deliberately constrain the decision space (available inputs or decisions) [13] or action space (available behaviours) [14], promoting increased stability for a particular autonomy context. Part of what still needs to be added is the appropriate language to underpin solution approaches as unified offerings for developing autonomous systems. 

Towards a coherent definition

At the start of this article, I provided an intuitive definition for MVA, given as 

The minimum amount of autonomous functionality required for a system to perform a task or set of tasks.

From our discussion, the essential aspect of a human custodian still needs to be added in the current definition for how or why a task should be performed and in what type of governance is required for operation. This requires a revised definition that balances the natural tension between system performance (effectiveness) and human expectations (reliability). 

Evolving the intuitive definition through the preceding enabling concepts definition, I offer the following concrete definition for MVA.

The minimum amount of autonomous capability that enables a system to complete its mission effectively and reliably. 

What's changed? 

This revised definition has three crucial changes from the intuitive definition, which are

• Including 'autonomous capability' instead of 'autonomous functionality'.

• Including 'complete its mission' instead of 'perform a task or set of tasks'.

• Including 'effectively and reliably'. 

The first two changes focus on autonomous capability. This broad term encompasses the holistic autonomous system's ability to achieve its mission instead of autonomous function, which focuses on more specific actions and tasks. Tightly coupled with the first change, the second change focuses on the performance of tasks and uplifts these to the more comprehensive requirement to complete a mission. Finally, including 'effectively and reliably' connects desired human benefits to increase teaming efficiency, underpinning the previously discussed context, predictability and stability concepts. 

Final thoughts

Minimum Viable Autonomy is an emerging concept for developing reliable, efficient, and safe autonomous systems. It ensures that these systems are well-engineered and tailored to meet specific operational needs effectively. As the field evolves, the definition and application of MVA will be refined, guiding the future of autonomous system development.

Often, increasing the level of autonomy in a system can be viewed as a mitigation of existing risks; however, this, too, can introduce novel risks that impact the promised potential of autonomous systems. An increase in autonomous complexity has the potential to iteratively introduce uncertainties, especially in the context of high-risk settings.

As the concept of MVA continues to evolve, it will become increasingly relevant in developing autonomous systems. The definition of MVA depends on the enabling concepts discussed as tenets guiding the approach. The goal remains to achieve the balanced integration of these elements, ensuring that autonomous systems enhance human capabilities without introducing undue complexity or uncertainty.

Not discussed in this article are the practical limitations and challenges of implementing MVA across various scenarios. While this article is theoretical and conceptual, real-world applications and further research are required to validate MVA empirically. Particularly, diving deeper into how predictability and stability are quantitatively measured for autonomous systems will illuminate further understanding of the practical implementation of MVA.

In this article, I've captured the essence of MVA by highlighting several important aspects that focus on functionality and utility, aligned with the broader idea that the design of autonomous systems should be just as complex as necessary but no more so. Such an idea has follow-on implications on autonomous systems resource optimisation, both at the design and implementation stages.

Join the discussion and share your thoughts on improving and implementing MVA for the real world.

Note: aspects of this article were written with the help of my research assistant, ChatGPT-4.

References

[1] Warwick, Graham. “Reliable Robotics Reveals Autonomous Caravan For Cargo” The Weekly of Business Aviation, 31 August 2020, https://aviationweek.com/sites/default/files/2020-08/bav_08_31_2020.pdf. Accessed 26 December 2023.

[2] Settimi, A., Pavan, C., Ferrati, M., Hoffman, E.M., Rocchi, A., Tsagarakis, N.G., & Bicchi, A. (2014). A Tele-Operation Tool for Humanoid Robots: On the Pilot Interface Design and Functionality.

[3] Ogor, P., Le Parc, P., Vareille, J., Marcé, L. (2001). Control a Robot on Internet,

IFAC Proceedings Volumes, Volume 34, Issue 29, Pages 103-108, https://doi.org/10.1016/S1474-6670(17)32801-X.

[4] R. Fernandez-Rojas, A. Perry, H. Singh, B. Campbell, S. Elsayed, R. Hunjet, and H. Abbass, ”Contextual Awareness in Human-Advanced-Vehicle Systems: A Survey," in IEEE Access, vol. 7, pp. 33304-33328, 2019, doi: 10.1109/ACCESS.2019.2902812.

[5] C.E. Shannon, “A Mathematical Theory of Communication,” The Bell System Technical Journal 27 (3): 379–423, Jul. 1948.

[6] A. J. Hepworth, D. P. Baxter, A. Hussein, K. J. Yaxley, E. Debie, and H. A. Abbass, “Human-swarm-teaming transparency and trust architecture,” IEEE/CAA J. Autom. Sinica, vol. 8, no. 7, pp. 1281–1295, Jul. 2021.

[7] ISO 31000 Risk Management, https://www.iso.org/iso-31000-risk-management.html

[8] Axelrod, R. (1990). The Concept of Stability in the Context of Conventional War in Europe. Journal of Peace Research, 27(3), 247–254. http://www.jstor.org/stable/423469

[9] Parliament. House of Lords (2023). Proceed with Caution: Artificial Intelligence in Weapon Systems. (HL 2023-2024 (16)). London: The Stationery Office. 

[10] Russell W. Glenn, "Mission Command in the Australian Army: A Contrast in Detail,” Parameters 47, no. 1 (2017), doi:10.55540/0031-1723.2833.

[11] Anna-Sophie Ulfert, Conny H. Antoni, and Thomas Ellwart. 2022. The role of agent autonomy in using decision support systems at work. Comput. Hum. Behav. 126, C (Jan 2022). https://doi.org/10.1016/j.chb.2021.106987

[12] Adriana Salatinoa, Arthur Prévelb, Émilie Casparc, & Salvatore Lo Buea. 2023. “Fire! Do Not Fire!”: A new paradigm testing how autonomous systems affect agency and moral decision-making. Pre-Print at bioRxiv. https://doi.org/10.1101/2023.12.19.572326

[13] S. J. Anderson, S. B. Karumanchi and K. Iagnemma, "Constraint-based planning and control for safe, semi-autonomous operation of vehicles," 2012 IEEE Intelligent Vehicles Symposium, Madrid, Spain, 2012, pp. 383-388, doi: 10.1109/IVS.2012.6232153.

[14] N. Mehr, R. Horowitz and A. D. Dragan, "Inferring and assisting with constraints in shared autonomy," 2016 IEEE 55th Conference on Decision and Control (CDC), Las Vegas, NV, USA, 2016, pp. 6689-6696, doi: 10.1109/CDC.2016.7799299.