PCBRouteAgent: How Close Are LLM Agents to Becoming PCB Routing Engineers?

Yue Zhang; Jilei Sun; Liqiang Jing; Yijie Wei; Ruohan Yang

cs.AINov 5, 2026

PCBRouteAgent: How Close Are LLM Agents to Becoming PCB Routing Engineers?

Authors: Yue Zhang, Jilei Sun, Liqiang Jing, Yijie Wei, Ruohan Yang

Organizations: The University of Texas at Dallas · Texas Instruments Inc. · Iowa State University

Abstract

Automating printed circuit board (PCB) routing requires coordinating board-level connectivity decisions with geometric execution under strict design rules. Although recent agents can interact with native routing tools, reliable routing on complex boards remains challenging. We present PCBRouteAgent, a multimodal PCB routing agent inspired by how human engineers coordinate global strategy with local execution. A Global Routing Planner combines board images and structured electrical and geometric information to specify connection topology, routing order, corridors, and layer transitions. A Detailed Routing Executor realizes these guides through KiCad's native routing engine and performs bounded local recovery. Unresolved failures are returned as structured feedback, enabling targeted guide revisions while preserving accepted routes. We evaluate PCBRouteAgent on 247 synthetic and real-world boards from four PCBWorld-Bench test splits. It achieves clean-pass rates of 100%, 94.9%, 100%, and 40% on D2, D3-A, D3-B, and D3-C, respectively. On the three real-world subsets, PCBRouteAgent outperforms the best evaluated baseline in clean-pass rate by 13.9, 36, and 10 percentage points, respectively.

Figures & tables

Figure 1: PCB routing connects required pads with copper traces and vias while satisfying design rules.

Method	CP $\uparrow$	Pot. $\uparrow$	Rout. $\uparrow$	DRV $\downarrow$	WL (mm) $\downarrow$	Via $\downarrow$
Freerouting	1.000	14.42	1.000	0.00	409.8	2.48
OrthoRoute	0.010	0.38	0.340	8.54	323.4	11.30
KiCadRoutingTools	1.000	15.52	1.000	0.00	373.9	8.09
PCBWorld (Haiku 5.5)	0.633	13.47	0.897	1.12	338.8	2.44
PCBWorld (GPT-5.4)	0.880	15.24	0.980	0.27	411.8	2.00
PCBWorld (GPT-5.4-mini)	0.200	6.46	0.490	4.62	208.6	0.96

Table 1: D2 routing quality (128 synthetic boards).

Method	CP $\uparrow$	Pot. $\uparrow$	Rout. $\uparrow$	DRV $\downarrow$	WL (mm) $\downarrow$	Via $\downarrow$
Reference	1.000	23.16	1.000	0.00	170.2	1.18
Freerouting	0.750	22.04	0.900	1.08	156.6	0.86
OrthoRoute	0.020	-6.16	0.530	99.70	129.8	26.97
KiCadRoutingTools	0.740	20.00	0.940	2.38	142.8	3.69
PCBWorld (Haiku 5.5)	0.343	14.77	0.721	4.06	111.5	0.96
PCBWorld (GPT-5.4)	0.420	16.15	0.810	2.71	148.1	2.36

Table 2: D3-A routing quality (99 real-world boards).

Method	CP $\uparrow$	Pot. $\uparrow$	Rout. $\uparrow$	DRV $\downarrow$	WL (mm) $\downarrow$	Via $\downarrow$
Reference	1.000	63.96	1.000	0.00	570.3	6.70
Freerouting	0.640	58.86	0.990	11.94	533.5	3.88
OrthoRoute	0.000	-10.25	0.440	188.20	370.0	39.80
KiCadRoutingTools	0.200	40.94	0.860	44.70	472.3	11.80
PCBWorld (Haiku 5.5)	0.000	25.32	0.442	25.40	245.7	1.90
PCBWorld (GPT-5.4)	0.000	25.65	0.440	24.88	240.3	2.06

Table 3: D3-B routing quality (10 real-world boards). Results include response-failure retries and angle repair.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Set	Method	CP $\uparrow$	Time (s) $\downarrow$	Calls $\downarrow$	In. (k) $\downarrow$	Out. (k) $\downarrow$	Cost ( $)$ \downarrow$
D2	PCBWorld	0.633	188.0	43.20	45.8	38.9	0.0240
(128)	PCBRouteAgent	1.000	0.8	0.05	1.6	0.1	0.0002
D3-A	PCBWorld	0.343	436.0	76.60	112.6	88.1	0.0553
(99)	PCBRouteAgent	0.929	27.0	0.87	19.9	5.8	0.0048
D3-B	PCBWorld	0.000	1817.0	263.90	900.0	352.0	0.2660
(10)	PCBRouteAgent	0.800	146.0	2.90	163.0	30.0	0.0330

Appendix

Table 5: Per-board routing success and computational cost with Claude Haiku 5.5 under the same evaluation conditions. Input and output tokens are in thousands (k). Within each split, bold and italic indicate the better and second-best values, respectively.

Set	Method	Connected edges	Tokens/edge $\downarrow$	$/clean-pass board$ \downarrow$
D2	PCBWorld	953/1096	11,374	0.0380
	PCBRouteAgent	1096/1096	202	0.0002
D3-A	PCBWorld	900/1301	22,079	0.1610
	PCBRouteAgent	1287/1301	1,972	0.0050
D3-B	PCBWorld	191/445	65,545	—
	PCBRouteAgent	443/445	4,355	0.0410

Appendix

Table 6: Resource use normalized by completed work with Claude Haiku 5.5. Connected edges are reported as completed/required totals for each split. Lower token and API costs are better.

Method	Calls	In. tokens (M)	Out. tokens (M)	API cost ($)	Time (h)
PCBWorld	22,529	80.3	25.4	20.72	36.4
PCBRouteAgent	172	10.5	1.5	4.77	2.6

Appendix

Table 7: Total computational cost across all 247 boards with Claude Haiku 5.5. Token counts are in millions (M); cumulative time sums the elapsed time of individual boards. Lower is better in every column.

Electronic Design Automation

Electronic Design Automation uses algorithms and software to design, verify, optimize, and manufacture electronic circuits and systems.

See all 227 papers

PCBWorld: A Benchmark Environment for Engine-Grounded PCB Design Automation
Jul 7, 2026Hyungseok Song, Junseok Park, Won-Seok Choi +4Electronic Design Automation AI Agent Benchmarks
LG AI Research Seoul, Republic of Korea
OmniRouting: A Semantic-Coupled Multimodal Benchmark for Constraint-Aware Spatial Reasoning in PCB Routing
Aug 5, 2026Taiting Lu, Kaiyuan Lin, Ziwei Dong +18Benchmark Design Electronic Design Automation
1Pennsylvania State University · 2Independent Researcher · 3Shanghai Jiao Tong University +3
pcbGPT: Automatic PCB Schematic Synthesis from Natural Language Requirements
May 31, 2026Tobias King, Steven Kehrberg, Michael Beigl +1Electronic Design Automation LLM-Based Program Synthesis
Karlsruhe Institute of Technology, Germany · Bosch Sensortec GmbH, Germany

Jul 7, 2026J/K move · Enter open · S save

PCBWorld: A Benchmark Environment for Engine-Grounded PCB Design Automation

Hyungseok Song, Junseok Park, Won-Seok Choi, Seohui Bae+3

LG AI Research Seoul, Republic of Korea

Electronic Design Automation AI Agent Benchmarks RL for Tool Use

Open paper AskPDF arXiv

PCB routing is the task of connecting the nets of a board with copper traces under strict design rules, yet learning-based methods still lag behind rule-based routers. We introduce PCBWorld, an open-source engine-grounded PCB routing environment built on KiCad, an electronic design automation (EDA) engine. As a human engineer does, agents in PCBWorld interactively route a board through the engine's native operations, guided by its Design Rule Check (DRC) feedback. The environment supports both RL and tool-using LLM agents. Alongside the environment, PCBWorld-Bench provides three board datasets in the native .kicad_pcb format, two controllable synthetic generators and 679 real open-source boards. It scores any completed board with eight engine-checked evaluation metrics, regardless of the routing method. In our experiments, agents in PCBWorld consistently outperformed grid-action RL policies and open-loop LLM baselines, and an RL policy trained only on synthetic boards transferred zero-shot to real boards, approaching rule-based routers. PCBWorld and PCBWorld-Bench are available at https://github.com/LGAI-Research/PCBWorld.

Abstract

Electronic Design Automation

Explore similar work

PCBWorld: A Benchmark Environment for Engine-Grounded PCB Design Automation

OmniRouting: A Semantic-Coupled Multimodal Benchmark for Constraint-Aware Spatial Reasoning in PCB Routing

pcbGPT: Automatic PCB Schematic Synthesis from Natural Language Requirements