Research Initiation: Manipulator Collision Avoidance and Dynamic Computation Using Neural Networks

该项目探索了神经网络的开发，用于执行复杂的计算：(A)在已知的静态工作环境下避免碰撞；(B)机械手动力学，即从关节位置、速度和加速度到关节扭矩/力的映射。对于碰撞避免，碰撞检测被看作是一个模式识别问题，可以用神经网络有效地、容易地解决。为了计算机械手动力学，正在研究一种三层神经网络来实现连续值输入和输出之间的映射。碰撞避免和机械手动力学都是计算非常密集的操作。对于实时控制使用，这些计算必须在可接受的时间段内完成，该时间段通常为几十毫秒或更短。神经网络使用大规模并行性，这使得这些计算成为可能。拟议的研究为解决这些问题提供了一种可行的方法。

The project explores the development of neural networks for performing complicated computations in (a) collision avoidance under known, static working environments, and (b) manipulator dynamics, which is a mapping from joint positions, velocities and accelerations to joint torques/forces. For collision avoidance, the detection of collisions is treated as a pattern recognition problem, which can be solved efficiently and easily by neural networks. For computing manipulator dynamics, a three-layer neural network is being studied for mapping between continuous- valued input and output. Collision avoidance and manipulator dynamics are very computationally intensive operations. For real-time control usage, these computations must be accomplished in an acceptable time period, which is usually tens of milliseconds or less. Neural networks use massive parallelism, which allows these computations to be possible in real time. The proposed research represents a feasible approach towards these problems.