Skip to main content
Back to timeline
arXivSource publication:

A machining agent stops computing its own toolpaths: MillBill holds 97.4% of target surface within 0.1 mm using geometric feedback

Synopsis

MillBill separates machining decisions from geometric execution: a locally hosted Qwen3.8-27B chooses operations, tools and cutting parameters, while a geometric harness generates toolpaths, simulates material removal and measures deviations. On 100 parts from GrabCAD, MillBill leaves 97.4% of the target surface within the 0.1 mm tolerance after simulated machining, against 94.3% for a GPT-6 Astra agent that writes its own geometry and toolpath code. The reference plan without the language model already reaches 97.0%, so most of the gain comes from the harness.

Source-provided article image: MillBill: A Machining Agent with Geometric Feedback
arXiv

Interpretation

Machining programming is split in two: the language model makes only process decisions, and coordinates, offsets and swept volumes are computed by geometric algorithms. Earlier language-model process planners output operation sequences, tools and cutting parameters but stop before toolpaths, so their plans are never executed against the part; other approaches have the model write G-code or its own geometry and toolpath code, where a single wrong number yields a program that runs but machines the wrong part. On 100 parts, MillBill with Qwen3.8-27B leaves 97.4% of the target surface within the 0.1 mm tolerance, against 94.3% for the code-writing GPT-6 Astra agent, 83.0% for the Opus 5.5 agent and 42.3% for CNC-Net.

The part is first segmented into machining features, and setups are chosen by what the cutters can actually reach rather than by visibility alone. Visibility is necessary but not sufficient for reach: a cutter has finite length and is clamped in a holder much wider than itself, so a deep, narrow region can be visible yet out of reach. The reach test runs one height-map dilation per remaining pair of holder overhang and reach over the whole tool library, then picks setups greedily by newly reachable volume; replacing reach-aware setups with a single setup from one direction drops the within-tolerance fraction from 97.4% to 66.4%.

The geometric harness turns the plan into toolpaths and G-code, tests material removal exactly against the analytically swept cutter, and attributes out-of-tolerance regions to features and operations for feedback. Common practice approximates simulation by discretising the workpiece with Z-maps, dexels or voxels, whereas here any query point is tested directly against the swept cutter. Each build is ranked by safety issues, gouged area and within-tolerance fraction, and the best of at most three builds is kept; one build gives 95.6%, a second raises it to 97.2% and a third to 97.4%.

Perspective

The evaluation covers 100 parts drawn from the GrabCAD Community Library at a 0.1 mm tolerance on an indexed three-axis machining centre, with one run per method and part; the same simulator scores MillBill and every baseline. The method is limited to three-axis machining from the six principal directions, and parts needing inclined setups or simultaneous five-axis motion are out of scope.

Workholding and machine travel are not modelled, toolpaths are not checked for collisions with fixtures or the tool holder, and the reach test is only a necessary condition for holder clearance; the simulation ignores tool deflection, vibration and wear, so a toolpath that meets tolerance in simulation may fail during machining. Taps and thread mills are not simulated, so thread geometry is planned but not verified; the knowledge base holds representative rather than vendor-specific cutting data, and AAGNet is trained on synthetic parts, so recognition errors can propagate into the plan.