Grafy News
Try Grafy free

MillBill Introduces Separate Decision and Geometry Layers for AI‑Driven Machining

A new AI agent called MillBill separates machining planning from geometric execution, improving reliability of CAD‑to‑toolpath conversion.

Illustration for the news story: MillBill Introduces Separate Decision and Geometry Layers for AI‑Driven MachiningSource: Grafy Research Lab

A new AI agent called MillBill can turn a CAD model into a three‑axis machining program while keeping geometric verification separate from the planning logic. The approach, described in a recent arXiv pre‑print, promises more reliable toolpath generation without overburdening a language model with low‑level calculations.

What is MillBill and why does it matter?

MillBill is an autonomous agent designed for indexed three‑axis milling. According to the authors, it receives a boundary‑representation (B‑Rep) of a part, the raw stock dimensions, and the material type, then proposes a sequence of machining operations together with cutting parameters. The novelty lies in delegating the actual generation and simulation of toolpaths to a dedicated geometric harness, rather than having the language model write raw G‑code itself. This separation mirrors how human machinists separate process planning from the precise numerical work of CAM software, potentially reducing errors that stem from language‑model hallucinations.

How does MillBill separate decision making from geometric execution?

The system consists of two loosely coupled modules. The first is a mid‑sized, locally hosted open‑source language model that reasons about manufacturing strategy: which features to face‑mill, the order of operations, spindle speeds, feed rates, and tool selections. The second module, referred to as the geometric harness, takes the textual plan and translates it into concrete toolpaths, runs simulations, and checks that the resulting geometry respects tolerances. By keeping the LLM out of the numerically intensive step, MillBill avoids having the model generate syntactically correct but physically impossible G‑code.

What components does MillBill rely on?

The paper lists three core ingredients:

  1. A boundary‑representation of the part – the exact CAD surface data that defines the shape to be machined.
  2. A locally hosted open model – the authors used a mid‑sized model that can run on commodity hardware, ensuring that the planning stage does not depend on external APIs.
  3. A geometric harness – a separate engine that builds toolpaths, runs a simulation of material removal, and verifies that the final geometry matches the target. The harness can also suggest adjustments if the simulated outcome deviates from the design. Together, these elements create a feedback loop: the LLM proposes, the harness validates, and the LLM can revise its plan based on the verification results.

How can developers experiment with a similar workflow in Grafy?

Grafy’s Code app provides a sandbox where you can assemble the two‑stage pipeline without leaving the browser. You could:

  • Load a B‑Rep file (e.g., STEP or IGES) using the Files tool.
  • Prompt a locally hosted open‑source model (such as Llama 2‑7B) via Chat to generate a high‑level machining plan.
  • Feed that plan into a custom script that calls an open‑source CAM library (e.g., PyCAM or OpenCAMLib) to produce toolpaths and run a simple simulation.
  • Use Graph to visualize the decision‑verification loop, branching whenever the harness reports a geometric mismatch. All of these steps stay within Grafy’s ecosystem, letting creators iterate quickly while keeping the heavy geometry work separate from the language‑model reasoning.

What are the current limitations and future directions?

The authors acknowledge that MillBill currently supports only indexed three‑axis milling, which covers many standard parts but excludes multi‑axis or high‑speed machining scenarios. The geometric harness also relies on classical simulation methods; integrating more advanced physics‑based models could improve accuracy for complex materials. Finally, the feedback loop is still linear – the LLM receives a single pass of verification before finalizing the plan. Future work might explore iterative reinforcement where the model continuously refines its strategy based on real‑time simulation data.

Where to learn more or try the code?

The full pre‑print, including eight figures and a link to the source code repository, is available on arXiv. Readers interested in reproducing the system can clone the repository, set up the LLM locally, and connect it to a CAM backend of their choice. For a hands‑on exploration within Grafy, start with the Code app and follow the workflow described above. The Reader app can also ingest the PDF and let you ask detailed questions about the methodology without leaving the platform.

Source: https://arxiv.org/abs/2611.22565

Frequently asked questions

How does MillBill ensure the generated toolpaths are geometrically accurate?

The geometric harness simulates material removal based on the LLM’s plan and checks the resulting shape against the original B‑Rep. If discrepancies exceed tolerance thresholds, the harness returns feedback so the LLM can adjust parameters before finalizing the G‑code.

Can MillBill be extended to multi‑axis machining?

The current implementation focuses on indexed three‑axis operations, but the authors suggest that the same decision‑verification architecture could be adapted to higher‑axis machines by expanding the geometric harness to handle additional degrees of freedom.

Do I need a powerful GPU to run MillBill’s language model?

MillBill uses a mid‑sized open model that runs on consumer‑grade hardware, such as a desktop GPU with 8‑12 GB of VRAM. This design choice keeps the system accessible without requiring cloud‑based inference.


Source: Grafy Research Lab — MillBill: A Machining Agent with Geometric Feedback