PCBench
A dataset of 164 KiCad boards with routing ground truth, augmentation scripts, and a reinforcement learning environment for measuring autorouter quality.
Summary rewritten from the vendor site because the catalog version did not match it.
What it does and how it connects
PCBench is a dataset for PCB routing, with 164 printed circuit boards stored as KiCad .kicad_pcb files. Each board folder holds the raw board, a cleaned-up board, a final.json with only routing-related information, a metadata.json, and a visualization. A master metadata file covers all boards and global-level information.
The README proposes a JSON specification called PCB Routing Description Language, or PCB-RDL, expressed in each board's final.json, to state a routing problem and its solution in basic concepts so machine learning research into automated routing can use it. The dataset also ships a data augmentation script that expands it for supervised learning tasks by randomly extracting a share of nets from a board.
A reinforcement learning environment is included for routing research. The dataset is aimed at measuring autorouter quality against real human routing, and it is the ground truth side of the freerouting self-improving loop in this batch: the loop scores optimizer changes against frozen baselines, and PCBench supplies real boards to compare against.
Alternatives
Compare side by side- DeepPCB
A cloud routing service for an already placed board, rather than a dataset of routed boards for measuring autorouter quality.
- freerouting self-improving loop
An agent-driven experiment loop inside the freerouting repo that scores routing optimizer changes against frozen baselines, rather than a dataset of routed boards.
- TraceMaker
Routes and places an existing KiCad .kicad_pcb file, rather than a dataset of routed boards.
- kicad-mcp (blwfish)
A community MCP server for KiCad workflows from description to manufacturing files, rather than a dataset of routed boards.
Recurring errors it addresses
Error Index entries this tool is built to reduce or work around.
- A KiCad MCP bridge normalizes DRC, connectivity, and geometry into its own output format, and that normalized output can silently disagree with the native KiCad engine. The maintainer had to build seeded false-PASS and false-FAIL tests because the bridge layer cannot be trusted as engineering truth on its own.Error Index
- KiCad lets a differential pair couple tighter than board clearance through a custom rule scoped to coupled pairs. This auto-router marks that rule unsupported and raises the pair gap to the net-class clearance instead; on thin dielectrics a 90 to 100 ohm pair needs a gap below board clearance, so the routed pair comes out at a different impedance than specified.Error Index
- An engineer running Claude Code against KiCad found the agent kept repeating a stale claim of 19 DRC violations that lived on in prose docs. MCP tool summaries cannot be treated as live numbers; only DRC message text and actual exports are authoritative.Error Index
- On six single edits to a public KiCad 10 design, three runs each, plain Claude Code with KiCad installed got 18 of 18 right; with the mcp-server-kicad plugin it got 3 of 18. The plugin had no tool for some edits, its rule against direct file edits made the model stop and ask, and its board-sync tool rewrote nearly every footprint. The maintainer put the project in maintenance mode on these results.Error Index
Sources
Not verified
- The directory record names the builder as the freerouting project; that was not confirmed in the sources read, and the repo org is PCBench, so the builder is recorded as PCBench.
- The directory record states the boards are converted to DSN fixtures; no DSN conversion was confirmed in the README as read.
- No GitHub releases exist, so latest release fields are null.
- No independent user voice found: recent issues are by the maintainer.
- Pricing: none stated on the repo.
- Activity: last push January 2024, so the dataset looks quiet; no newer release was found.
Last verified November 6, 2026. Listings are informational. The evidence score measures public, verifiable evidence and is not an endorsement or a compatibility claim.