Soft Value Iteration Networks for Planetary Rover Path Planning

Value iteration networks are an approximation of the value iteration (VI) algorithm implemented with convolutional neural networks to make VI fully differentiable. In this work, we study these networks in the context of robot motion planning, with a focus on applications to planetary rovers. The key challenging task in learning-based motion planning is to learn a transformation from terrain observations to a suitable navigation reward function. In order to deal with complex terrain observations and policy learning, we propose a value iteration recurrence, referred to as the soft value iteration network (SVIN). SVIN is designed to produce more effective training gradients through the value iteration network. It relies on a soft policy model, where the policy is represented with a probability distribution over all possible actions, rather than a deterministic policy that returns only the best action. We demonstrate the effectiveness of the proposed method in robot motion planning scenarios. In particular, we study the application of SVIN to very challenging problems in planetary rover navigation and present early training results on data gathered by the Curiosity rover that is currently operating on Mars.