2. Implementation of a RAS to provide the required behaviour

Contextual description: Having defined how the RAS must behave in order to be sufficiently safe, it is then necessary to implement the RAS such that it provides that behaviour throughout its life, and to provide sufficient evidence that this has been achieved. In order to define this appropriately for an RAS, there are a number of objectives that must be satisfied, as described below. System requirements are implemented through a process of architecture and design decomposition. Although this process may vary enormously for different systems and domains, it is generally possible to consider an RAS in terms of an agent model consisting of the following elements:

• Sensing
• Understanding
• Deciding
• Acting

The relationship between these elements is indicated in Figure 1 below. Each of these elements may be further decomposed into components that implement that aspect of the RAS behaviour. Note that not all components need be part of the RAS itself — they may be part of infrastructure provided externally, e.g. an autonomous car may perform some Sensing by receiving information from roadside beacons.).

Next sections:

• 2.1 System-level verification (guidance available)
• 2.2 Implementation of SUDA elements (guidance available)
• 2.3 Implementing requirements using machine learning (guidance available)
• 2.4 Controlling interactions with other systems (guidance available)
• 2.5 Controlling interactions at the system level
• 2.6 Handling change during operation (guidance available)
• 2.7 Using simulation (guidance available)
• 2.8 Explainability (guidance available)