Embodied CAD: Solver-Grounded LLM Agents for Parametric B-Rep Assembly Modeling
Embodied CAD integrates LLM planning with geometric solvers via a hierarchical skill library, enabling reliable parametric assembly modeling.
Key Findings
Methodology
This approach combines LLM-based high-level planning with deterministic modules for parameter resolution and CAD execution, forming a closed-loop system. The framework employs a stratified L0-L4 skill hierarchy, where the LLM predicts operation families, which are resolved into specific geometric parameters via deterministic functions. These parameters are executed in a CAD backend, with solver feedback guiding corrections and learning. The training employs supervised fine-tuning and GRPO-style reinforcement, leveraging solver diagnostics such as success, topology, and volume checks. Experiments demonstrate high task success rates, with deterministic workflows achieving 100% execution success and learned controllers reaching over 93% exact policy accuracy across multi-category industrial assembly tasks.
Key Results
- In multi-step mechanical, industrial, and mold assembly tasks, deterministic planners achieved 100% execution success, while learned controllers attained over 93% exact policy accuracy. The skill prediction model reached 76.6% overall accuracy under boundary, random-parameter, and paraphrase stress tests. Solver-guided planning ensures all steps satisfy geometric constraints, significantly outperforming sequence-based methods. Ablation studies confirm the importance of operation family prediction and geometric validation, with the framework effectively handling complex long-horizon workflows.
- Solver feedback enables the model to detect and repair geometric errors, improving editability and robustness. The GRPO reinforcement strategy enhances decision-making by providing structured reward signals, reducing errors in long sequences. Comparative analysis shows that models with explicit geometric validation outperform purely visual or heuristic baselines, highlighting the critical role of solver integration in industrial CAD automation.
- Ablation results indicate that operation family prediction reduces index sensitivity and improves accuracy, while hierarchical skill libraries constrain action spaces for better planning reliability. The system demonstrates strong generalization across diverse assembly tasks, with high success rates and repair capabilities, validating the effectiveness of the closed-loop, solver-grounded approach for industrial applications.
Significance
This work advances the automation of industrial CAD assembly by embedding geometric validation within a learning framework, addressing longstanding issues of error propagation and geometric inconsistency in long workflows. It bridges the gap between AI planning and precise geometric execution, enabling autonomous, editable, and reliable parametric assembly. The approach enhances manufacturing efficiency, reduces manual intervention, and paves the way for intelligent robotic assembly systems. By integrating solver feedback into the learning process, the method ensures geometric correctness and robustness, crucial for real-world industrial deployment. The framework’s success demonstrates the potential for AI-driven design automation, fostering deeper integration of AI and CAD in manufacturing and engineering.
Technical Contribution
The paper introduces a hierarchical skill-based planning framework that tightly couples LLMs with deterministic geometric modules and solvers. It innovates by predicting operation families instead of instance-specific commands, reducing index sensitivity. The deterministic resolver translates high-level predictions into precise parameters, while solver feedback provides structured rewards for reinforcement learning. This combination ensures geometric validity at each step, enabling long-horizon, multi-part assembly workflows. The framework also employs a multi-level skill hierarchy (L0-L4) to constrain action space and improve planning reliability. These contributions collectively enable robust, editable, and scalable parametric CAD assembly automation, representing a significant leap over sequence-only generation methods.
Novelty
This work is the first to embed a solver-grounded, hierarchical skill library within an LLM-based planning framework for industrial CAD assembly. Unlike prior sequence generation models, it emphasizes operation family prediction and geometric validation, ensuring each step’s correctness. The integration of solver feedback as a structured reward signal for reinforcement learning is novel, effectively bridging high-level planning with low-level geometric execution. The hierarchical skill design reduces index sensitivity and enhances generalization, addressing key limitations of previous methods. Overall, this approach marks a new paradigm in AI-driven parametric CAD modeling, combining planning, validation, and learning in a unified framework.
Limitations
- The current system relies heavily on predefined skill hierarchies and operation families, limiting flexibility for non-mechanical or highly organic shapes, which require new skills and validation checks.
- Geometric validation and solver computations are computationally intensive, affecting real-time performance and scalability in large or complex assemblies.
- Error accumulation over long workflows persists, especially under extreme geometric or boundary conditions, indicating the need for more robust error correction mechanisms.
Future Work
Future research will focus on expanding the skill library to cover more diverse design domains, including organic shapes and free-form surfaces. Enhancing solver efficiency and integrating multi-modal perception (visual, tactile) will improve robustness and applicability. Developing adaptive skill learning and transfer mechanisms could enable rapid domain adaptation. Additionally, incorporating uncertainty modeling and active repair strategies will further strengthen long-horizon planning robustness, pushing towards fully autonomous, real-time industrial CAD assembly systems.
AI Executive Summary
Deep Dive
Plain Language Accessible to non-experts
想象你在一个工厂里，工人们需要组装一台复杂的机器。每个步骤都像拼拼图一样，有很多零件要放对位置、钻孔、切割，然后组合起来。传统的方法就像让工人记住一长串指令，但很容易出错，比如放错零件或位置不对。现在，这个新系统像是给工人配备了一个聪明的助手，它不仅能理解每个步骤，还能在装配过程中不断检查，确保每个零件都放得正确，甚至能自己修正错误。这个助手会先想好每个动作，然后在实际操作中不断确认和调整，确保最后的机器完美无缺。它就像一个既会计划又会自己检查的机器人，能帮工厂更快、更准地完成装配任务。这让工业生产变得更智能、更可靠，也让未来的制造变得更高效。
ELI14 Explained like you're 14
想象你在玩拼装玩具，但这个拼装机器人比你还聪明！它知道每个零件应该放在哪里，怎么拼，甚至能自己检查拼错了没有。如果它放错了一个零件，它会马上发现，自己修正，然后继续拼。这个机器人用一种特别聪明的方法，先想好每一步，然后一步步去做，还会用“检查器”确认每个步骤都正确。这样，拼装出来的机械就不会出错，也不用你一直盯着它看。它就像一个超级聪明的拼装助手，既会计划，也会自己修正错误，最终拼出一台完美的机器。这个技术让工厂里的机器人变得更聪明、更可靠，也让我们的工业生产更快更好！
Abstract
Large language models can write plausible CAD scripts, but reliable industrial CAD modeling requires more than syntactically valid code: every feature, placement, and assembly relation must be accepted by an exact geometric kernel while remaining editable as parametric boundary representation geometry. We present Embodied CAD, solver-grounded LLM agents for parametric B-Rep assembly modeling. Instead of generating a complete script in one pass, the agent iteratively selects actions from a stratified L0-L4 CAD skill library, resolves them into typed geometric operations, executes them in a CAD backend, and uses solver feedback to plan, repair, and learn. The framework combines action grammar constraints, deterministic parameter resolution, and solver-derived rewards for supervised warm-up and GRPO-style refinement. We evaluate Embodied CAD on multi-step mechanical, industrial equipment, and mold-oriented assembly tasks using solver-aligned metrics: executable rate, skill accuracy, operation-family accuracy, exact policy accuracy, and task completion success. The results show that solver-grounded planning executes all strong-planner workflows in the current benchmark, while learned controllers reach high executable rates and expose the remaining gap between valid tool calls and exact long-horizon policy prediction.