Abstract:
In this proposal, I address the problem of autonomous high-speed navigation for rally racing and military applications. State-of-the-art methods in autonomous navigation are inefficient in their local planning and perception methods, limiting their capabilities to drive fast. Inspired by research in cognitive science and experimental psychology, I study the main computational benefit of allocentric representations and exploit their advantages to propose an efficient local navigation system. I propose an allocentric planning framework that exploits motion prediction in multiple allocentric frames to reduce unnecessary and expensive re-planning computations. I show that the proposed planning framework could be supported directly and efficiently by a robust and fast dual-vision system, mimicking the selectivity mechanism of bio-vision for efficient resource allocations. As a result, the system can fluidly coordinate local perception and planning together. I propose to implement and evaluate the entire system on a high-end car simulator, X-Motor Racing, as well as on a real RC car platform.