Embedded Decision in the LAAS Architecture

Despite the presence of an AI planner (STRIPS) on one of the very first mobile robot (Shakey), one must admit that few of our nowadays robots embed and run a tasks/activities planner. However, the need for embedding such a deliberative activity is definitely present. For example, if one consider exploration robotics, the ESA Exomars project (part of Aurora) aims at having the rover, during one day, navigating over its “field of view” with navigation and activity decisions taken on board. NASA MSL will also push the autonomy cursor further than for MER1. Last, the “Human on Mars” goal will require the deployment of a large number of autonomous systems to “prepare” and study the planet before a human can set foot on it. It is clear that the “future” of exploration rovers and probes lies in an increased autonomy addressing the problems of action planning, and plan execution control. Even closer to us, service robots will need a high level planner, to provide the robustness one expects. Meanwhile, recent developments in the AI planning community, combined with the computational power available on robots now allow for embedding planning activities. Indeed, planning systems are now taking into account time and resources, which are two essential features for planning in the real world. Moreover, they are “execution” aware and either produce: contingent plans, plans as policies, plans dealing with uncertainties, or as we will show in this paper plans which can be repaired (locally) while other parts remain executable. If one looks at the current state of the art, few high level planning systems have been integrated onboard real robots, many architectures (such as Claraty [3]) provide a “decisional” level where such components can sit, but little was done as far as deploying it on real systems. The main reason is probably that despite the availability of good planning systems, few of them integrate the proper plan repair and replanning mechanisms. Still, the ROGUE system [4], for instance, performs planning for asynchronous goals and execution monitoring enhanced with learning capabilities. In [5], the authors propose a different approach where the plans themselves specify the adaptation processes as subplans. In any case, very few