A Neuromorphic Controller with On-Chip Learning for Robot Motion Control

Motion control is an important issue in robotics, which encompasses two typical control scenarios, repetitive point-to-point and dynamic random trajectory motion control. This brief proposes a dual-mode neuromorphic controller with on-chip learning. It can be configured in Adaptive Control Model mode for repetitive point-to-point motion control and Spiking-Proportional-Integral-Derivative (S-PID) mode for random trajectory motion control. When operating ACM, the controller implements on-chip learning through the Prescribed Error Sensitivity (PES) rule, thereby enabling real-time synaptic weight updating via continuous error signal feedback. Thus, the controller can dynamically compensate for exogenous disturbances through learning, it achieves sub-millimeter accuracy and millisecond latency with external disturbances, yielding 2.9× and 2.0× enhancement of control precision and convergence speed over conventional approach. When performing S-PID, through event-driven error regulation and temporal spike encoding mechanisms inherent to the Spiking Neural Network (SNN), the Root Mean Squared Error (RMSE) and the Integrated Time Absolute Error (ITAE) of dynamic arbitrary trajectory motion are improved by at least 1.3× and 1.5×, respectively. The controller is simulated under 40-nm CMOS technology, occupying a core area of 0.93 mm2 and consuming 3.4 mW of power at a frequency of 50 MHz, satisfying the requirements of high control performance and low cost for different robot motion control tasks.