Executive Summary : | The robots are now being widely used to solve complex problems that are hard to be learnt in a straight-forward mechanism. Specifically considering the problem of a robot navigating amidst humans, there are several complex situations that the robot may enter into that are hard to learn. The proposal builds upon the unique expertise developed at IIIT Allahabad in handling robot navigation in cluttered environments through planning and reactive approaches along with their general expertise in learning for vision along with the expertise at the Bielefeld University to handle complex learning problems especially the ones involving the use of simulations and reinforcement learning for complex environments. Together the two groups understand their complementary strengths and propose to develop a technology that make robots and virtual robotic agents in simulation learn in complex environments for complex tasks, using planned paths in a latent space as an enabler. The aim is to integrate planning in a learning framework to simply a complex problem that can be learnt. The approach proposes to gather multiple motion sequences of a robot performing complex tasks, and to cluster the space into a low dimensional latent encoding, to derive a planning problem that can be used to give suitable sub-goals to the robot. The problem then asks the robot to achieve suitable sub-goals that is generally easier to learn. The learning in such a way also learns to extract the contextual information for its operation and use the same for decision making. Using this technology, IIIT Allahabad shall specifically develop social robotics solutions, while the Bielefeld University proposes to use the same technology for solving open challenges and games using simulations. |