Creative Mobile Terrain Sensing Multi-Valued Behavior Robots

In previous investigations, we have crafted neural architectures for complex, hexapod robots that may invent clever walking strategies based upon absolutely no prior learning. Similarly, we have also shown that the same neural networks may devise variations on these locomotive strategies to propel themselves along unfamiliar surfaces. In this SBIR, we propose using similar techniques to teach a hexapod robot, equipped with a more extensive sensor suite, to not only create its own crawling strategies for non-ideal surfaces such as grassy, sandy, wet, or snowy terrains, but to associate signatures from its sensor suite with these self-invented strategies. In operation, the robot will thereby enlist the appropriate crawling behaviors stored within corresponding neural engines. Having selected an appropriate crawling behavior, the neural algorithm will further refine itself as it continuously experiments and absorbs the most successful strategies through reinforcement learning. In the final stages of this SBIR, we will devise a series of test courses over a variety of surfaces and relief features to test our advanced, neural network based navigation algorithms.