Hive Mind: A Framework for Problem Solving in Multi-Agent Systems

This project focuses on developing artificial intelligence algorithms for learning physics models from observations of both computational simulation and real-world data. The final aim is for a robot to use these models to perform basic tasks such as grasping and picking an object with a manipulator. The research hypothesis is that by allowing the robot to learn to predict the effect of its behaviour, we can achieve better training and runtime performance when executing tasks.
The research will be carried out, in a first stage, by developing reinforcement learning techniques (i.e. training a robot to achieve a goal through trial and error) in a graphics and physics simulation engine. Data captured with real depth camera sensors will be fed back into the simulation to refine environment variables, potentially leading to better behaviour models.
While object grasping has been subject of significant research over the past years with already successful applications in industry, the incorporation of physics models will expand these techniques to work with a wider range of objects, taking into account additional features such as perceived weight and friction.