Skip to content

AutoCAM Pipeline

Home / Engineering / System Design / CAM Pipeline / AutoCAM Pipeline

Status: Skeleton

This document is a starting point. Sections marked with TBD need to be filled in.

Purpose

The AutoCAM Pipeline is the core of the pipeline and the highest-risk component. It takes recognized features and produces complete, validated candidate manufacturing plans — including orientations, fixturing, toolpaths, and operation sequences — optimized for our standardized cell.

Candidate Generation Pipeline

AutoCAM generates multiple candidate manufacturing plans in parallel and ranks them to select the best option. The pipeline fans out at two key split points — orientation selection and toolpath strategy — then merges results into a ranked set of feasible plans.

flowchart LR
    FEAT["Feature\nRecognition\n(shared)"]
    FEAT --> SPLIT1{{"Split:\nOrientations"}}

    subgraph PER_ORI ["× N orientations"]
        direction LR
        SETUP["Fixturing &\nStock Selection"]
        SETUP --> SPLIT2{{"Split:\nStrategies"}}

        subgraph PER_STRAT ["× M strategies"]
            direction LR
            OP1["Op 1: Toolpaths\n& Sim"]
            OP1 --> SJ["Soft-Jaw\nGen"]
            OP1 --> OP1CUT["Op 1\nCutting"]
            SJ --> OP2["Op 2: Toolpaths\n& Sim"]
            OP1CUT --> OP2
        end

        SPLIT2 --> OP1
    end

    SPLIT1 --> SETUP
    OP2 --> MERGE{{"Merge &\nRank"}}
    MERGE --> OUT["Feasible Plans\n(ranked)"]

Step 1: Feature Recognition (Shared)

Feature recognition runs once on the input part geometry and produces a structured feature map. This output is shared across all candidates — every orientation and toolpath strategy works from the same feature data.

Step 2: Candidate Orientations (First Split Point)

Generate a set of candidate part orientations to explore. Each orientation defines how the part sits in the vise for Op 1 and, by extension, what geometry is accessible in each setup. This is the first fan-out point in the pipeline — the number of orientations directly multiplies the downstream work, so balancing coverage against compute cost is important.

A pre-processing step may be needed to align parts into logical orientations before candidate generation. Many CAD files arrive with arbitrary coordinate systems, so normalizing to a principal-axis alignment (e.g., longest dimension along X) gives orientation strategies a consistent starting point.

Potential orientation strategies:

  • Brute force — enumerate all 6 axis-aligned faces (±X, ±Y, ±Z), filter symmetrical duplicates. Simple and exhaustive for axis-aligned parts.
  • Feature-driven — analyze the feature map to identify orientations that maximize feature accessibility per setup (e.g., prefer orientations where the most holes and pockets are reachable from the top).
  • Learned — train a model on historical parts and their successful orientations to predict high-probability candidates, reducing the search space.

Initial strategy: Brute Force

Start with the brute force approach — try all 6 axis-aligned orientations, then filter out:

  • Symmetrical duplicates — orientations that produce equivalent setups due to part symmetry
  • Physical infeasibility — orientations where the part + fixture doesn't fit within the DVF 5000 work envelope

This assumes most parts are designed in axis-aligned ways, which holds for the majority of prismatic aluminum parts we expect from target customers. More sophisticated strategies can layer on top as we learn which orientations succeed most often.

Step 3: Per-Candidate Planning

For each candidate orientation, the following stages run independently:

Fixturing & Stock Selection

Fixturing: Standardized dovetail vise (Lang Makro-Grip on zero-point). AutoCAM selects vise size based on billet dimensions. Fixture models are included in all downstream simulation.

Stock selection: Determine the stock (billet) for this orientation. Two approaches:

  • Auto-generate from the part's bounding box plus a machining buffer
  • Select from standard stock sizes maintained in the system's inventory

Stock choice affects cycle time (more material to remove = longer cuts) and may constrain which orientations are viable.

Op 1: Toolpath Generation, Probing & Simulation

Select tools from the fixed 120-tool HSK-A63 library and generate collision-free toolpaths for Op 1. Assign operations (roughing, finishing, drilling, thread milling) and determine execution order.

  • Built on a proven CAM kernel (ModuleWorks)
  • Thread milling (Emuge Thriller + Threads-All) instead of tapping — no tapping chucks, salvageable parts on tool break
  • Feeds and speeds optimized for 6061-T651 aluminum

Probing routines are generated alongside toolpaths and included in the NC program:

  • Datum probing (before cutting) — locates the workpiece in the vise so all cuts reference the correct coordinate system. Uses the Renishaw OMP40-2/OMP60 spindle probe to find part edges, faces, or reference features.
  • End-of-op probing (after cutting completes for the op) — verifies critical features to confirm the op went as expected before moving on. This on-machine check catches gross errors; more careful inspection happens off-machine in the metrology lab.

Mid-cycle probing between individual cuts (e.g., probing a bore after semi-finish to adjust the finish pass) is a future optimization, not part of the initial approach.

Toolpaths and probing moves are then validated together with fast stock removal simulation (similar to CNCSim from ModuleWorks) — must run in seconds per plan, not minutes. Candidates are rejected based on:

  • Collision detection — tool, holder, probe, or spindle collisions with the part, fixture, or machine (applies to both cutting and probing moves)
  • Kinematic feasibility — moves that exceed axis limits or approach singularities (e.g., gimbal lock in the B/C rotary axes)

Toolpath Strategies (Second Split Point)

Within each orientation, different toolpath algorithms or strategies can be applied, creating a second fan-out point. For example, different roughing approaches (adaptive vs. conventional), operation sequences, or tool selection strategies can be explored in parallel to increase the chances of finding a good plan.

This is where the × M strategies multiplier in the diagram comes from — each orientation may produce multiple candidate plans, one per strategy. All are evaluated through stock removal simulation and only feasible plans survive to the merge step.

  • TBD: Which strategies to implement first and how to define the strategy interface
  • TBD: RL optimization approach for learning which strategies work best on which geometry classes

Soft-Jaw Generation (Parallel)

While Op 1 is being cut on the physical machine, soft-jaw geometry for Op 2 is generated and sent to the SLS printer (Formlabs Fuse). This runs in parallel so that fixtures are ready by the time Op 1 completes.

Op 2: Toolpath Generation, Probing & Simulation

After the simulated part flip, the same process repeats for Op 2: toolpath generation, probing routine generation (datum probing to locate the part in the soft jaws, end-of-op verification), and stock removal simulation — all using the soft-jaw fixture and the intermediate stock state from Op 1. The same rejection criteria apply — collision detection and kinematic feasibility for both cutting and probing moves.

Step 4: Merge & Rank

All surviving candidates merge into a single ranked list. The ranking heuristic considers:

  • Cycle time — estimated total machining time across both ops
  • Tool count — fewer tool changes is preferred
  • Number of setups — simpler plans are preferred when quality is comparable

The top candidates proceed to full machine simulation for detailed validation.

See Machine Simulation & Validation for the thorough outer-loop simulation that runs after this stage.

Interfaces

Direction System Data
Input Feature Recognition Structured feature map
Input Standardized tool library 120-tool definitions (geometry, holders, reach)
Input Fixture library Dovetail vise models (3 sizes), soft-jaw generation parameters
Output Machine Simulation 3–5 candidate plans (NC programs with interleaved probing routines, tool lists, fixture models, work offsets, intermediate stock states)
Output Operations Integration Soft-jaw geometry for 3D printing, stock size request
Feedback Machine Integration Real cutting data for RL training and model calibration

Open Questions

  • What is the RL training loop architecture? How does real machine feedback update the model?
  • How do we measure and track automation coverage (% of parts that need no human CAM intervention)?
  • What is the fallback workflow when AutoCAM cannot produce a valid plan? (Currently: manual programming in Mastercam by CMO)