Collision-free　navigation　is　an　important　research　direction　for　multi-robot　systems,　in　which　the　two　core　problems　are　navigating　to　the　target　point　and　avoiding　other　robots.　Many　researchers　use　deep　reinforcement　learning　as　navigation　strategy　to　realize　multi-robot　collision　avoidance　navigation.　However,　most　of　them　use　raw　sensor　information　or　global　state　information　of　the　agent　as　the　neural　network　input,　which　is　not　conducive　to　extending　the　navigation　strategy　to　a　larger　space.　This　paper　proposes　an　improved　deep　reinforcement　learning　navigation　strategy,　which　enables　robots　to　learn　navigation　and　collision　avoidance　strategies　more　accurately.　This　strategy　converts　the　interactive　environment　state　from　the　global　coordinate　representation　to　the　relative　vector　representation,　and　attenuates　the　influence　of　the　rear　irrelevant　agents　on　the　collision　avoidance　strategy.　Experimental　results　show　that　the　proposed　method　outperforms　existing　learning-based　methods　in　three　indicators:　success　rate,　additional　time　to　reach　the　target,　and　model　convergence　speed.　©　2023　SPIE.