Exploiting Probabilistic Knowledge under Uncertain Sensing for Efficient Robot Behaviour
Marc Hanheide, Nick Hawes, Charles Gretton, Alper Aydemir, Hendrik Zender, Andrzej Pronobis, Jeremy Wyatt and Moritz Göbelbecker
One of the most significant challenges in AI is to develop robot systems that exhibit intelligent goal-directed behaviour in real world environments. Achieving this means overcoming problems due to uncertain perception and action, whilst also exploiting the cues and structure present in the world. Towards meeting this challenge, we have developed an integrated robot system that is able to opportunistically build and exploit rich probabilistic models of uncertainty while performing tasks. Our system includes a novel spatial representation based on a probabilistic graphical model. This provides a mechanism to integrate probabilistic ontological knowledge about the world with instance knowledge derived from noisy sensors. We generate system behaviour based on this representation using a unique continual planning system that synthesises efficient plans by automatically switching between probabilistic and deterministic approaches. We have evaluated our system on an object search task in two different indoor environments. The results demonstrate the concrete benefits arising from the integration of probabilistic mechanisms into our system.