- CAD
- Read, edit and correct engineering drawings; model parts from their drawings; assemble products from their STEP files. Drawings come as PNG for vision agents and as DXF for agents that use tools.
- ECAD
- Read, change and correct circuit-board designs in KiCad, across schematic, board and settings. Built on public open-source boards, with the source and the licence in every task.
- CAE
- Set up, run and check simulations: find the loads and supports from the geometry, mesh, solve, check convergence and report the quantities the brief asks for.
- Engineering data
- The drawing sets, 3D models, solver decks and manufacturing files behind the tasks, also on their own.
Environments
CAD, CAE and CAM environments for engineering agents.
Talos builds RL environments from complete engineering data. In every task an agent gets drawings, 3D models and a specification, works with open-source engineering tools in a sandboxed terminal, and submits files that a program scores. We build each environment end to end, from the data to the verifier, and deliver it ready to run in your harness.
What we build
Full pipeline
One design, from drawing to finished part.
We cover the whole engineering chain: design, simulation and manufacturing. A design can be followed from its drawing to its simulation to the finished part or board, and its data stays consistent at every stage. Tasks can cross the stages too, for example checking that a drawing can be machined as dimensioned, or that a board fits its machined enclosure.
- 01 · CAD · ECADDesignDrawing sets (DWG, DXF), 3D models (STEP with PMI, native CAD), circuit-board projects.
- 02 · CAESimulationLoad cases and materials, meshes, solver input decks, reference results.
- 03 · CAMManufacturingCNC programs and process plans; PCB fabrication and assembly files (Gerber, drill, BOM, pick-and-place).
Partner CNC, PCB and PCBA factories
Difficulty
Four tiers, set by the type of work, the same for CAD and ECAD. Entry tasks are a warm-up; task families that score above their target are made harder.
- Entry
- Read and reason about one drawing or one board: values, fits and tolerances, what follows from the standards.
- Core
- Find and fix errors across a drawing set or a board; apply a change order everywhere it applies; assemble a sub-assembly.
- Hard
- Model a part from its drawing; assemble a whole product; a structural analysis from geometry alone.
- Flagship
- Model and assemble a complete product in one run; the dynamics of a whole mechanism.
How tasks are scored
- A program scores every task, never an LLM judge. The reward is continuous, from 0 to 1.
- The verifier runs in its own container, without network access, with pinned tool versions.
- Before delivery, the reference solution scores 1.0 and known wrong submissions score below 0.2.
- A submission that times out or runs out of memory scores 0, so a policy cannot escape a low reward by breaking the verifier.
- Every task design ships seeded variants with their own answers, so a policy cannot memorise them. Held-out sets are split by design.
Data quality
- Original
- Every CAD and CAE design is checked for public availability; designs found online are not used. The ECAD line is the stated exception.
- Consistent
- Parts lists, title blocks and 3D models are cross-checked, and every part position in an assembly is verified face by face.
- Clean
- English drawings in ISO terminology, with no company names, authors, timestamps or software traces.
- Complete
- Everything needed to solve a task is in the task. Difficulty never comes from missing or ambiguous information.
What a task ships with
Tasks ship in the Harbor format, so they run in any harness that reads it, and rebuild byte for byte.
- instruction.md
- The statement: inputs, the work, deliverables with exact keys and units, and the scoring rule.
- task.toml
- Tier, time-outs, CPU and memory limits, no network for the agent or the verifier, a separate verifier.
- environment/
- The agent's image: pinned open-source tools such as OpenCASCADE, ezdxf, gmsh and CalculiX, read-only inputs, a non-root user.
- solution/
- The reference solution. It scores 1.0.
- tests/
- The verifier's image. It writes the reward and an item-by-item score with its evidence.
- qc.json
- The self-check record: the reference score, the scores of known wrong submissions, the byte-identical rebuild.
Access
- Evaluate
- Run your model on a held-out set under the same conditions. Held-out tasks are never licensed for training. Ask about evaluation
- Train
- Environments and engineering data for post-training and RL, in the Harbor format. Get a sample