Neural network approaches to real-time motion planning and control of robotic systems

Real-time motion planning and control are fundamentally important in robotics. In this thesis, a framework, based on biologically inspired neural networks, is developed for real-time robot motion planning with obstacle avoidance in a nonstationary environment. Each neuron in the topologically organized neural network is characterized by a shunting equation or an additive equation. The developed algorithms can be applied to point mobile robots, manipulation robots, holonomic and nonholonomic robots, and multi-robot systems. The planned real-time robot motion with safety consideration does not suffer from either the “too far” or the “too close” problems. The real-time optimal robot motion is planned through the dynamic activity landscape of the neural network without explicitly searching over the free workspace or the collision paths, without explicitly optimizing any cost functions, without any prior knowledge of the dynamic environment, without any learning procedures, and without any local collision checking procedures at each step of robot movement. Therefore the proposed algorithms are computationally efficient. The computational complexity linearly depends on the neural network size. The global stability and convergence of the neural network system is guaranteed by both qualitative analysis and the Lyapunov stability theory. The model algorithms are not sensitive to model parameter variations nor sensor noise. The last part of the thesis presents an efficient neural network based approach to real-time fine motion control of robot manipulators with completely unknown robot dynamics and subject to significant uncertainties. The real-time fine robot motion control is achieved through only the on-line learning of the neural network, without any off-line training procedures. The proposed controller is capable of quickly compensating sudden changes in the robot dynamics. The neural network assumes a single-layer structure, and the learning algorithm is computationally efficient. The global asymptotic stability of the system and the convergence of the tracking error is proved by the Lyapunov stability theory.