Assembly skill acquisition via reinforcement learning

A neural network controller is proposed for the motion control of robot manipulators with force/torque feedback signals. This controller is trained with reinforcement learning algorithms and a model is extracted from the synaptic weights within the neural network. This model is continuously refined by the feedback signals to ensure its validity even in a stochastic and non‐stationary environment. With this model and the real‐time force/torque feedback data, the robot can acquire a fine skill for a particular assembly task for which it is trained.