Isometric SolidWorks view of the Revision D arm prototype on a white background, showing the yaw platform, both pitch joints and the modular parallel gripper on its bench base plate
SolidWorks 2026 · Revision D assembly
03 · Personal Project · Jun 2026 – Present

EOAT Automation Cell — 4-DOF Arm Prototype

In progress · Design phase
4
Actuated DOF — yaw, shoulder, elbow, grip
198
Components in the native assembly
15
Printed part types, 23 instances
32 mm
Parallel jaw stroke (purchased tool)
SolidWorks 2026SolidWorks API / Python Design for 3D PrintingTolerance & Fit Servo SizingCOTS IntegrationParallel GripperAllen-Bradley Micro820 Arduino / C++Test Planning
⏳
Nothing on this page has been built or measured yet. The arm exists as a fully constrained native SolidWorks assembly with working joints, now at Revision D; every number below is a CAD-derived estimate or a design target, and is labelled as such. No parts have been printed and no hardware has been purchased. The Results table stays empty until it is filled from physical test data.
View the repo on GitHub
The gap I wanted to close

On a production floor I worked on tooling that a robot and a PLC already drove — I owned the mechanical end of the problem, not the sequencing, sensing, or fault handling around it. This project is the deliberate inverse: a cell small enough that I own every layer of it, from the locating scheme on the part to the state machine that decides when the gripper is allowed to close.

The whole cell is built around one pick object — an 8 × 22 × 7 mm 608ZZ bearing — located by an ID pin and rest at the pickup pose and a different datum scheme at the destination. Fixing the part first, then designing everything else to serve it, is the same order of operations that industrial fixturing follows, and it is what keeps the scope from drifting.

The single statement that decides whether this project worked:

Success criterion

The system shall autonomously detect, pick, transfer, and place a part for 50 consecutive cycles with at least 98% successful completion, documented fault handling, and measured positional repeatability.

The prototype is a 4-DOF articulated arm: base yaw, shoulder pitch, elbow pitch, and a gripper. It is modelled natively in SolidWorks 2026 as editable parts and a real mated assembly — not a static arrangement of meshes — so joint travel can actually be exercised in CAD.

Revision D is the fourth pass, and it changed the architecture rather than the details. The printed single-finger gripper is gone, replaced by a purchased parallel gripper on a removable printed fork, so the tool is now a swappable interface instead of part of the arm. Both pitch joints moved to MG996R servos sharing one common horn interface — the same OD20 / PCD14 metal horn and 32 mm standoff spacing at the shoulder and the elbow — and the old single yaw-platform-plus-upright part was split into a separate platform and a bolt-on upright so each one prints and services independently.

Joint Architecture
Servo horn takes torque only
608ZZ bearings take radial load
6.35 × 25.4 mm steel dowels as pivots
Printed dowel-to-8 mm-bore adapters
Screwed keepers for axial retention
Serviceability
Removable yaw servo cassette
Yaw platform and upright are separate bolt-on parts
Gripper fork unbolts as a unit (4 × M3×16)
One horn interface serves both pitch joints
Crush spacers stop the tool ears clamping inward
Sizing Basis
20 g payload target (provisional)
Solid-density mass bound, 1.25 load factor
Budget = ⅓ of reference stall torque
Shoulder margin 1.15× — the binding constraint
Elbow margin 3.60× on the second MG996R
0.65 N per jaw needed at µ = 0.3
Reference ratings, not measured torque
The problem, as it stood at Revision B

The shoulder was modelled against a nominal 26 mm horn plate on a 20 mm bolt circle. The metal disc horn actually sold for this servo class is 20 mm OD on a 14 mm bolt circle — it would not drop into the modelled pattern, and substituting it also shifted the axial stack by about 2.5 mm. Every horn-connected part was held from release, and three different horn patterns across four servos meant three separate interfaces to reconcile.

How Revision D closed it

Both pitch joints were standardised onto the real OD20 / PCD14 M3-tapped metal horn — the part that is actually sold — at 20.5 mm spacing on 32 mm standoffs, and the printed links were rebuilt around it rather than around a nominal envelope. Dropping the printed gripper removed a fourth horn interface entirely. What was three mismatched patterns is now one measured pattern used twice, and the supplier drawing is saved next to the model in source-reference/.

This is worth stating plainly because it is the whole argument for modelling properly before printing: the mismatch was found by a CAD check, cost four revisions of rework, and cost no material and no money. Finding it with a printed link and a bag of the wrong screws is the alternative. Seated horn height and the bearing/dowel fits are still assumptions — those need calipers on the real parts, not another revision.

Shoulder margin · 1.15×

Two MG996R cases and a 30 g tool sit outboard on a 150 mm lever, and modelled mass is up to 637 g. Against a budget of one third of reference stall torque the shoulder now has 1.15× where Revision B had 2.15×. That is a CAD estimate built on a solid-density mass bound and a datasheet rating, so the true figure could fall either side of it — which is exactly why it has to be measured on the built arm before anything downstream is trusted.

Tool availability

The gripper is a $29.95 candidate that the supplier showed as backordered on 2026-09-12, and nothing has been purchased. The fork adapter is built around its published STEP and dimension drawing rather than a measured unit, so the interface is verified against the supplier's geometry, not against hardware in hand.

Four phases, each independently demonstrable. No phase starts before the previous one has measured results — the failure mode for a project like this is fragmenting it into five half-finished subsystems.

PhaseGoalDemonstrable outputStatus
1 — Mechanical proof Printed arm, one pick object, defined pickup and place poses, repeatable hard stops Manual cycle with basic repeatability data CAD in progress
2 — Motion & sensing Arduino motion control, home switch, part-present sensor, end-position confirmation Repeatable automatic pick-and-place cycle Not started
3 — PLC supervisory control Micro820 owns cycle start, permissives, state sequence, faults, reset and the handshake; the motion controller owns low-level movement Fault injected, detected, reported, recovered Not started
4 — Validation Repeatability, pick rate, cycle time, fault recovery, payload, gripper retention, sensor robustness Completed test report Not started

The PLC-to-motion-controller split is deliberate. Letting the PLC supervise a separate motion controller over a simple I/O handshake is how a real cell is partitioned, and it is the part of automation I had the least ownership of on the floor.

Empty until measured. This table gets filled from the test plan, not from expectations.

MetricTargetMeasured
Successful pick rate≥ 98% over 50 cycles—
Positional repeatabilitySet after Phase 1—
Cycle timeSet after Phase 2—
Payload≥ 20 g—
Gripper retention through transferNo drops across the validation run—
Lock the part before you design the machine
Choosing the 608ZZ first — and writing down how it gets located at both ends of the cycle — is what made every later decision decidable. Jaw geometry, payload, Z clearance over the pin, and the sensor choice all fall out of that one commitment. The scope only started converging once the part stopped being a variable.
A model that can't move hasn't been checked
Building real mates instead of a fixed arrangement is slower, and it is what caught the pickup pose colliding with the yaw platform. Driving the joints through named angle dimensions and then independently comparing all 198 component transforms against the kinematics turns "it looks fine" into a check that either passes or does not.
Four revisions to admit the right part was for sale
Revisions A through C all had a printed single-finger gripper, and each one made it slightly better without making it good. Swapping to a $29.95 parallel gripper solved repeatability, stroke and feedback in one move, and turned the interesting problem into an interface problem — mounting, crush spacers, fastener access — which is the problem worth having. Knowing which parts to design and which to buy is a judgement call, and I got it wrong for three passes.
Every improvement spends something
The second MG996R and the purchased tool made the arm more capable and more serviceable. They also added 119 g, moved mass outboard, cut the shoulder's usable range from −30° to −15°, and halved its torque margin. Nothing about Revision D is free, and writing down what it cost is the only way the next revision starts from the truth rather than from the highlight reel.
Separate what you assumed from what you measured
The load report labels every input: CAD solid volume, manufacturer reference mass, assumed friction coefficient. The torque margins are real conclusions from those inputs, but they are not test data, and writing them down that way keeps me from quoting a 1.15× margin as if a servo had actually been loaded. The same discipline is why the Results table above is still blank.
Optimus Academy All projects Towed-Sensor Winch