PCBRouteAgent coordinates plans with physical PCB routing
PCBRouteAgent: How Close Are LLM Agents to Becoming PCB Routing Engineers? · arXiv:2611.08175
In plain words
A circuit board needs copper paths joining the right component pads, while leaving enough room for other paths and following design rules. PCBRouteAgent first plans which connections to make and where they might go, then asks KiCad’s routing engine to draw the actual traces. If a route gets stuck, it can revise its plan.
On 99 small real-world test boards, it passed every connection and basic rule check on 94.9% of boards, compared with 81.0% for the strongest tested baseline. Large boards remained harder: six of ten were incomplete.
The research
A completed connection can make the next one harder by using scarce board space.
Results on real designs can change expectations about whether a routing agent should plan across a board and revisit guides when routes fail, while the unfinished large-board cases show that this coordination still has limits.
PCBRouteAgent pairs board-level routing plans with KiCad’s physical router and revises guides from failures, improving clean completion on tested real-world boards.
In PCBWorld-Bench, each board’s reference routes were removed while its placement, netlist, and design rules stayed fixed. KiCad’s official evaluator counted a clean pass only when all required connections were complete with zero error-level violations.
On 99 small real-world boards, PCBRouteAgent passed 94.9%, versus 81.0% for the strongest evaluated baseline. On 10 large boards, six remained incomplete, bounding the gains on harder designs.
The results cover fixed benchmark splits, and the D3-B result includes retries of model-response failures and angle repair. One case study also shows that complete connectivity and basic rule checks can miss signal-specific electrical requirements.
What the evidence shows
- Plan across the board
PCBRouteAgent is a multimodal PCB routing agent that coordinates board-level plans with physical trace routing. Its planner uses board images and structured electrical and geometric details to choose connection groups, routing order, corridors, and candidate layer changes.
KiCad’s native router then lays the traces and vias. This division matters because a route can be locally feasible yet consume space needed for other connections.
- Use failures to revise guides
The system’s feedback loop addresses cases where a guide meets obstacles the planner could not fully anticipate. The executor checks whether pad groups actually became connected and tries bounded local recovery when routing fails.
If a connection remains unresolved, the planner receives structured feedback and can revise its guide while accepted routes remain in place. The contribution is coordination between planning and execution, rather than exact trace geometry from the planner.
- Compare complete boards
PCBWorld-Bench tests whether routing systems can complete board connections while meeting basic design rules. A clean pass requires every connection to be complete with no error-level violations.
On 99 small real-world boards, PCBRouteAgent’s clean-pass rate was 94.9%, compared with 81.0% for the strongest evaluated baseline. This comparison shows an advantage on that test split.
- Large boards remain difficult
On ten large real-world boards, PCBRouteAgent had a higher clean-pass rate than Freerouting, but most of its boards still did not pass.
This gap shows why average connection progress cannot stand in for completing a whole board: a board passes cleanly only when every required connection is complete and there are no error-level design-rule violations.
- Revisions help, with limits
A control that kept the detailed executor and scripted local recovery but removed language-model guide revisions had no clean passes on the medium or large boards. This comparison supports a contribution from guide revisions within this framework.
It does not establish how they compare with a dedicated rule-based replanner or a general-purpose agent using the same execution tools. A case study also shows a boundary: connectivity and basic rule checks can miss signal-specific electrical requirements, so electrical validation is still needed.
Deep
- Original paper authors
- Yue Zhang, Jilei Sun, Liqiang Jing, Yijie Wei, Ruohan Yang
- Submitted to arXiv
- License
- arXiv non-exclusive