The last decade has seen a rapid increase in the number and diversity of deployed autonomous systems, including AGVs, but also those striving for full autonomous operation in unstructured environments, for example in automotive and aviation applications. Enabled by technological advances in sensors, computing, and especially AI, all of these novel systems face a challenge of demonstrating consistently safe behavior in even limited deployments, much less at the scale needed for future operations.
An advantage of AGVs in characterizing safe operation is the ability to better define the environment in which it operates, eliminating many of the concerns of operations in unstructured domains. Even so, demonstrating safe operation of limited numbers of AGVs in highly controlled environments is difficult enough; characterizing safety in future systems where increasing levels of autonomy enabled by AI and broader operational domains limit the control of infrastructure is even more daunting.
The question for AGV operators, as with all autonomous vehicles attempting to deploy at scale, is how to verify safe operation of complex, novel systems, particularly in a world lacking comprehensive prescriptive measures for safety evaluation?
Fundamentally, any safety assessment will attempt to characterize the likelihood of unsafe interactions and the consequence of those should they occur. Controlling the operational domain is arguably the most effective means to reduce both, but increasing scale (and potentially autonomy) always introduces uncertainties and risk that must be characterized and either reduced or mitigated. That can often largely be accomplished with careful simulation followed by testing and validation.
Simulation capabilities have improved significantly with emerging techniques in machine learning and AI. However, simulation will continue to be necessary but insufficient in the development and validation of complex systems, particularly as those systems increase in scale. This necessitates additional testing to verify operational performance and safety as well as to validate the simulation models themselves.
As scale increases, so does the number and complexity of interactions between AGVs and the objects around it, and maintaining a consistent structured environment becomes increasingly difficult. As the complexity (and uncertainty) increases, a highly robust system is forced to become more resilient and the requirement for state knowledge becomes paramount. In practical terms, for AGVs, perhaps the most important state measure involves an understanding of the physical position and potential action of the vehicle with respect to the objects around it.
Measuring system state is, of course, complex on its own, especially when addressing safety, which by its very nature requires accurate knowledge of the external environment. A large variety of sophisticated sensing modalities are often used, including video, laser, and lidar systems, which can characterize the position of AGVs with respect to their environment with great accuracy. Recent advances in object detection also allow for highly accurate identification and detailed characterization of a large variety of human and non-human objects along with their associated positions and movements (tracks). This, in turn, enables an understanding of the kinematic relationship of these objects relative to the AGV that provides an important hazard metric, enabling a better understanding of the likelihood and consequence of physical contact.
But how are these and similar assessments used to demonstrate safe operation for scaled systems? From a process perspective, there is an increasing reliance on development of structured, but still largely subjective, Safety Cases used to provide evidence of safe operation for autonomous systems. Originally developed in Europe, the Safety Case is a set of highly structured arguments supported by relevant evidence demonstrating safe operation of a system in a particular environment. They are often deployed as a key element for those systems too complex, or too novel, to have associated prescriptive standards or regulations. This often occurs when rapidly scaling existing systems in ways that exacerbate and magnify hazards in unexpected ways or when actual incidents or injuries are too infrequent to provide meaningful statistical evidence.
By the end of this decade, higher-performance position and object detection/identification enabled by continuing advances in both sensing and AI will begin to blur the performance differences between autonomous guided and fully autonomous operation. AGV systems with both increased sensing performance and greater intelligence will tend to automate more functionality while being demonstratively safer and, with less need to rely on structured environments, likely cheaper and more resilient. The ability to better characterize the system state, in turn, allows for more predictable, controlled scaling while ensuring safety is not compromised.
Ultimately, the realization of this future is dependent on the ability to understand the state of the system and its surroundings in sufficient detail to be able to predict the likelihood of physical interactions between the system and, as it pertains to safety, the humans within its operational domain. These advances, coupled with increasing AI-enabled intelligence, are inexorable, and will undoubtedly improve the performance of AGV systems at scale, while doing so at lower cost. More importantly, they will allow for safer, more resilient operation at the same time.
