PurposeThis　study　aims　to　address　the　issue　that　existing　methods　for　limb　action　recognition　typically　assume　a　fixed　wearing　orientation　of　inertial　sensors,　which　is　not　the　case　in　real-world　human-robot　interaction　due　to　variations　in　how　operators　wear　it,　installation　errors,　and　sensor　movement　during　operation.Design/methodology/approachTo　address　the　resulting　decrease　in　recognition　accuracy,　this　paper　introduced　a　data　transformation　algorithm　that　integrated　the　Euclidean　norm　with　singular　value　decomposition.　This　algorithm　effectively　mitigates　the　impact　of　orientation　errors　on　data　collected　by　inertial　sensors.　To　further　enhance　recognition　accuracy,　this　paper　proposed　a　method　for　extracting　features　that　incorporate　both　time-domain　and　time-frequency　domain　features,　markedly　improving　the　algorithm＇s　robustness.　This　paper　used　five　classifiers　to　conduct　comparative　experiments　on　action　recognition.　Finally,　this　paper　built　an　experimental　human-robot　interaction　platform.FindingsThe　experimental　results　demonstrate　that　the　proposed　method　achieved　an　average　action　recognition　accuracy　of　96.4%,　conclusively　proving　its　effectiveness.　This　approach　allows　for　the　recognition　of　data　from　sensors　placed　in　any　orientation,　using　only　training　samples　conducted　at　an　orientation.Originality/valueThis　study　addresses　the　challenge　of　reduced　accuracy　in　limb　action　recognition　caused　by　sensor　misorientation.　The　human-robot　interaction　system　developed　in　this　paper　was　experimentally　verified　to　effectively　and　efficiently　guide　the　industrial　robot　to　perform　tasks　based　on　the　operator＇s　limb　actions.