We propose a quasi-layered recurrent neural network consisting of sensing neurons(upper layer)and driving neurons(lower layer). In both layers, chaotic dynamics are used where, in sensing neurons, sensitive response to external input is utilized, whereas in driving neurons, complex dynamics is utilized to generate complex motions. These two properties are applied to solving two-dimensional mazes by computer simulations and hardware implementation into a roving robot is shown.Keywords: roving robot · chaotic dynamics · quasi-layered RNN · complex control · adaptive control · ill-posed problem PR0001/07/0000-0976 ¥400 © 2007 SICE