A　method　of　realizing　desktop　robot＇s　negative　phototaxis　through　a　neural　network　is　presented.　The　biology　is　characteristic　of　biologic　phototaxis　and　negative　phototaxis.　Can　a　machine　be　endowed　with　such　a　characteristic?　This　is　the　question　we　study　in　this　paper.　A　randomly　generated　network　is　used　as　the　main　computational　unit.　Only　the　weights　of　the　output　units　from　this　network　are　adjusted　during　the　training　phase.　Learning　samples　are　collected　according　to　the　energy　function.　It　will　be　shown　that　this　simple　type　of　a　biological　realistic　neural　network　is　able　to　simulate　robot　controllers　like　that　incorporated　in　desktop　robots.　The　experiments　are　presented　illustrating　the　stage-like　study　emerging　with　this　learning　mode.