In this paper, an iterative learning control approach, clean system inversion, is proposed for multi-input multi-output (MIMO) timeinvariant systems, as a remedy to the system inversion law. The analysis and design are carried out in the frequency domain. This model based scheme relies on non-causal filtering and need no high order derivatives of error signals and no numerical differentiation. Experimental results on an industrial robot system show that the scheme is effective. (c) 2005 Elsevier Ltd. All rights reserved.