Neural network path planning applied to PUMA560 robot arm

Abstract

Abstract: In this paper, we present a trajectory planning method using a recurrent neural network with strictly limited interconnections. The trajectory obtained is then applied to PUMA560 robot arm which moves in an environment with obstacles. Each neuron of the network is connected only to the nearer neighboring neurons. This makes possible to reduce considerably the number of interconnections and to thus decrease the complexity of the network. The neural network evolves from an initial state to a final state, thus delivering an optimal trajectory that the robot must follow to avoid the obstacles and to reach the desired configuration. In order to avoid local minima, we use an adaptive parameter in the neural network activity equation. The most significant feature in this method is that it can be established to have a real-time trajectory planning in a dynamic environment. Key-Words: Recurrent neural network, Real-Time-Cost, Robot navigation, Collision avoidance.