In　recent　years,　we　have　witnessed　the　rapid　development　of　the　Internet　of　Things　(IoT)　paradigm.　A　new　computing　paradigm,　predictive　intelligence,　has　been　introduced　into　many　IoT　applications.　In　IoT　applications,　predictive　intelligent　devices　must　consider　power　and　delay　consumption.　The　goal　of　Very　Large　Scale　Integration　(VLSI)　circuit　routing　problem　is　to　minimize　the　wirelength　to　optimize　chip　delay.　Therefore,　VLSI　routing　problems　have　been　widely　applied　in　IoT.　The　X-architecture　routing　technology　presents　a　promising　solution　that　can　reduce　the　number　of　vias,　wirelength,　and　mold　size　significantly,　thereby　proving　to　be　a　highly　practical　approach.　Moreover,　with　the　shift　of　modern　VLSI　design　towards　system-on-chip　paradigm,　many　uncrossable　obstacles　appear　on　the　plane,　thus　Obstacle-Avoiding　X-architecture　Steiner　Minimal　Tree　(OAXSMT)　problem　has　attracted　more　and　more　attention.　Most　existing　algorithms　are　troubled　by　poor　wirelength　quality　or　high　construction　time　cost.　Therefore,　this　paper　proposes　a　method　based　on　Deep　Reinforcement　Learning　to　construct　an　OAXSMT.　First,　an　encoder　based　on　heterogeneous　graph　transformer　is　designed　to　generate　fixed-length　encodings　for　each　node.　Second,　an　actor-critic　decoder　based　on　pointer　network　is　designed　to　generate　obstacle-avoiding　X-architecture　Steiner　redundant　tree.　Finally,　a　fast　evaluation　strategy　is　designed　to　evaluate　the　quality　of　the　redundant　tree.　Experimental　results　show　that　the　routing　results　generated　by　the　algorithm　are　superior　to　existing　heuristic　algorithms　in　terms　of　wirelength　and　runtime,　and　the　algorithm　also　has　certain　generalization　ability.　©　2023　IEEE.