A recent study that evaluated issues associated with remote interaction with an autonomous vehicle within the framework of grounding found that missing contextual information led to uncertainty in the interpretation of collected data, and so introduced errors into the command logic of the vehicle. As the vehicles became more autonomous through the activation of additional capabilities, more errors were made. This is an inefficient use of the platform, since the behavior of remotely located autonomous vehicles didn’t coincide with the “mental models” of human operators.
One of the conclusions of the study was that there should be a way for the autonomous vehicles to describe what action they choose and why. Robotic agents with enough self-awareness to dynamically adjust the information conveyed back to the Operations Center based on a detail level component analysis of requests could provide this description capability. One way to accomplish this is to map the behavior base of the robot into a formal mathematical framework called a cost-calculus. A cost-calculus uses composition operators to build up sequences of behaviors that can then be compared to what is observed using well-known inference mechanisms.
The explanation system is broken up into three subsystems that address the principal developments needed:
1. An inference mechanism for the mapping of observed behaviors into the cost-calculus: The observation equivalence of behaviors on a single autonomous agent and between two or more agents is done through bi-simulation relations. An example of the inference mechanism at work in a Rules-of-the-Road behavior is shown in the figure.
2. A learning mechanism for the cost-expression generation for observed behaviors outside of the cost-calculus tactical behavior base: Reinforcement learning of observed behavior patterns is used for the common grounding of behaviors sequences that were not previously observed, or that are in the command dictionary of the autonomous agent.
3. Explanation capabilities for the system: A dynamic decision tree decomposition of the observed behaviors is used to generate a set of rules for explanation. An adaptive level of detail is automatically built into this process in that all of the sensory information that led to a behavior is available, and can be conveyed to the operator if the human/machine interface (HMI) has a detail level of request capability.