AI agents can now iterate through mechanical CAD assemblies step by step, catching geometric errors in real time rather than at the end of a failed script.
Researchers have published Embodied CAD, a system that pairs large language models with an exact geometric solver to handle parametric boundary representation modeling. Rather than generating a complete CAD script in one shot, the agent picks actions from a tiered skill library, resolves them into typed geometric operations, executes them in a CAD backend, and uses solver feedback to correct mistakes and refine future moves. The training pipeline combines grammar constraints and solver-derived rewards, including a reinforcement learning method called GRPO, to push the agent toward geometrically valid, editable assemblies.
This matters because the failure mode of existing LLM-generated CAD code is silent: the script runs, but the resulting geometry is broken or non-parametric, making it useless in real engineering workflows. By treating solver acceptance as the ground truth rather than syntactic validity, Embodied CAD reframes the problem around what industrial CAD actually requires. Results on mechanical, industrial equipment, and mold assembly benchmarks show solver-grounded planning completing all tested high-difficulty workflows, though a gap remains between valid tool calls and correct long-horizon policy prediction.
CAD automation has attracted steady research attention, but most prior work targets script generation without closing the loop on geometric correctness — the same gap that has kept AI-assisted design mostly out of production engineering toolchains.