Using recurrent neural networks (RNNs) as planners for bio-inspired robotic motion

In this paper, we propose using Long Short Term Memory Networks (LSTM) to serve as planners for bio-inspired robotic motion. LSTMs can learn long or short term correlations of sequential data. Using LSTM networks, we implement a motion planner using simulated fish trajectories. The motion or path planning unit can then be implemented on robots such that they can operate autonomously without knowing their absolute position in a global frame. Simulation results show that the planned path demonstrates characteristics that are similar to simulated fish trajectories. This work may lead to learning animal behavior and then formulating bio-inspired path planners for robots to operate in unknown environments.