Controlling A Redundant Articulated Robot In Task Space With Spiking Neurons

Using spiking neural network (SNN) controllers [3] to implement multiple motor tasks for complex redundant robots requires efficient methods to compute complex kinematic and dynamic functions with spiking neurons. Three fundamental problems arise while using SNNs to compute high-dimensional robot kinematics using steady-state spike rate decoding (following the neural engineering framework [2]): first, differential maps from the generalized coordinates to task-space, task control Jacobians, cease to be humanly factorizable into subfunctions with low-dimensional domains; second, efficient Jacobian factorizations require multiple neuron layers, exacerbating neuron spike noise and latency; and third, function-agnostic sampling strategies require an exponential growth in the number of neural response samples as the number of input dimensions increases. Here, we present an SNN implementation that overcomes these problems to compute kinematic functions (Jacobians) for the Kuka LBR iiwa, and Kinova JACO, which have seven and six degrees-of-freedom respectively. Both robots are redundant for task space motion control. To control them, we developed a multi-task control system where task Jacobians, and a part of the Jacobian's dynamically consistent generalized inverse, were implemented with SNNs. Our SNN was an asynchronous spiking neural simulation with dynamical neurons modeled using the Neurogrid neuromorphic system's soma equations [1]; it thus serves as a model of what neuromorphic computers can achieve.