Learning-Based Adaptive Optimal Tracking Control for Flexible-Joint Robots With Quantized States: Theory and Experiment

This study aims to overcome a problem in the position tracking control of flexible-joint robots: realizing good position tracking for desired signal while conserving bandwidth and minimizing cost. The primary obstacle is that the weight update laws developed through the reinforcement learning (RL) scheme fail to guarantee a bounded quantized signal. Hence, an optimal controller is designed based on the bounded effect of the proposed fuzzy basis function, with the signal discontinuity problem caused by the quantized virtual controller addressed via command filter technique. Meanwhile, an adaptive law is designed to replace the model identification, allowing it to handle unknown structure impacts and reduce approximation behaviors. Besides, we establish an improved compensation signal to maintain boundedness via a first-order low-pass filter. Notably, the developed scheme guarantees boundedness of all signals. Finally, the justification of the proposed scheme can be further confirmed by the simulation and the comparison experiment on Quanser hardware experiment platform, which shows that the developed technology can achieve desired tracking performance.