Adaptive teleoperation using neural network-based predictive control

Teleoperation systems strive to accurately render often unstructured environments to operators. However, due to the existing delays in the communication channel, transparent performance and stability are compromised. This paper presents a new class of teleoperation predictive controllers, in which the dynamics of the environment is mapped and simulated at the master side using two neural networks. The supervised network at the slave side is trained online to generate environment contact force using slave contact position and force. The master network whose gains are adaptively updated online with the transmitted slave network gains, replicate the environment force using master position. The estimated environment force is utilized in a "pseudo" two-channel force-position bilateral teleoperation control architecture. The proposed controller does not require an environment model to reflect environment dynamics for transparency. Thus, it can be used for operations on unstructured environments displaying varying nonlinear dynamic behavior. The improved performance of the new teleoperation architecture in comparison with that of a conventional two-channel force-position architecture that uses measured environment force for feedback is verified on a teleoperation test-bed consisting of two planar Twin-Pantograph haptic devices