Engineering Process

Design decisions are only credible if the corrections are visible too.

This page documents how Subway Sentinel was actually engineered: what was tried, what failed, what the measurement said, and what changed because of it. The project moved from an early motion concept into a full Fusion 360 mechanical layout, then into structural screening, mass-property measurement, and a logged RobotStudio inspection cycle.

Every controlling dimension traces to a vendor drawing, a published source, or a measurement taken inside the CAD model. Where a value is a proxy or an assumption, it is labeled as one. That labeling discipline keeps the analysis results later in this page traceable and auditable at concept level.

The tunnel came first. Everything else was sized to fit inside it.

Subway Sentinel is built around the IRT Broadway–Seventh Avenue Line, the 1 line, in the Washington Heights deep-tunnel section near 168th, 181st, and 191st Streets.[10]

Research into the West Side Line showed the 1 line is not one tunnel type. It combines cut-and-cover subway, viaduct, and deep rock tunnel construction built by different contractors and methods. Modeling one generic tunnel would have produced a design that fits nowhere. The CAD baseline was narrowed to a typical two-track deep rock tunnel, roughly 25 feet wide by 18 feet high with 12-foot track centers, taken from historical IRT engineering references.[9]

Track gauge is modeled at standard gauge, 1,435 mm or 4 feet 8.5 inches, measured between the inside faces of the rail heads rather than between rail centerlines. The New York City Subway runs standard gauge across all divisions, the same as most North American railroads.[11]

Rail section took more digging. The current MTA qualified products list maintains fastener hardware for two running-rail sections in parallel, each with its own NYCT standard drawing series: 100-8, the modern designation for the 100 lb ARA Type B section the original IRT was built with, and the heavier 115RE.[12] Stocked quantities for 100-8 hardware run several times higher than 115RE, which indicates 100-8 remains the dominant installed section while 115RE appears in newer and heavier work. This project models 115RE with 1:40 inward cant as a modern concept baseline, drawn directly in Fusion rather than imported so gauge, cant, and rail-head geometry stay under project control. It is a design selection, not a claim that any specific location on the line carries that section.

Train clearance uses current MTA A-Division vehicle envelope dimensions from the open-gangway design requirements rather than a guessed car size.[8]

A scaled R62A source mesh provides visual scale only. Its dimensions were checked against a published R62A dimension sheet, then it was converted to SAT and STEP solely as RobotStudio scenery. It is excluded from collision and clearance acceptance; the simplified R262/A-Division envelope solid controls the static vehicle-clearance check.[8][13][14]

A frame that resists twisting, not just weight.

The rail platform is the core mechanical contribution. It carries the 1,780 kg ABB IRB 6750S-215/3.9, its OmniCore V250XT controller, the deck, and the inspection payload, and it has to stay stiff while the robot reaches sideways toward a tunnel wall.

The first structural decision was to stop trying to design a railcar. A full rail vehicle would require suspension, wheelset, braking, and certification work far outside this project's scope. Building on an existing bogie geometry kept the engineering effort on the parts that are actually new: the platform frame, the robot mounting structure, the tooling, and the validation chain.

Because the dominant load is torsional rather than purely vertical, the frame uses closed rectangular sections instead of open beams. Closed sections carry torsion far better, which is well documented in chassis torsional stiffness studies.[15][16] Every structural member was sketched and dimensioned rather than approximated.

The robot mount is a structure, not a plate.

The ABB baseplate bolt pattern and fastener specification were taken directly from the robot manual.[2] From field experience, industrial robots at this size are normally raised on a pedestal rather than bolted flat to a deck, so the mount was developed as a welded torque box between the baseplate and the frame. The internal ribbing uses a star pattern to hold stiffness while removing mass, and gusseted ribs carry the load out into the bogie interface adapter.

The deck is 30 removable panels of 8 mm tread plate, measured in CAD at 1,875 kg. That mass is not incidental. It sits low and centered, and adding it is what moved the worst-case stability result above the project threshold.

Every interface was dimensioned from the manufacturer's own drawing.

The ABB IRB 6750S-215/3.9 was selected for reach and payload after motion studies, then integrated through a tool changer, a custom wrist adapter, a generative-design bracket, and a carrier for the ground-penetrating radar.[1]

The ATI QC-160 tool changer was chosen partly from prior field experience with ATI hardware and partly through the manufacturer's Quick Consult selection tool, then confirmed against the official dimensioned drawing.[3][4] That confirmation mattered. The vendor selection guidance implied a direct mount, but the drawings show the ABB wrist uses a 100 mm H7 locating feature on a 160 mm bolt circle with eleven M12 fasteners, while the QC-160 customer interface uses an 80 mm H7 feature on a 125 mm bolt circle with ten M10 fasteners. Those patterns are not interchangeable, so a custom adapter plate was required: a 30 mm aluminum 7075 plate carrying both patterns, measured at 2.48 kg, with a flange-to-coupling stack of 85.423 mm. The adapter is concept geometry; no structural analysis is claimed for it.

The bracket was designed twice.

The bracket connecting the tool changer to the radar carrier was developed with Fusion's generative design solver. The first attempt was rejected. It used a generic dynamic load factor rather than published robot data, its moment axes did not rotate with the applied force, it had no load case for the tool hanging on its stand, and the side plate was too thin for the ATI hook adapter.

The study was rebuilt on ABB's published maximum emergency-stop acceleration of 45.3 m/s², applied in three orthogonal directions with the corresponding center-of-gravity offset moment rotated per case, plus a fourth load case representing the tool seated on the stand.[1] Two materials were run in parallel. Aluminum 7075 and 6061 produced effectively identical geometry, so 6061 was selected: lighter in this case, easier to machine, and consistent with the rest of the carrier.

The radar payload is modeled on the IDS GeoRadar Stream T in its full six-antenna configuration, a 35 kg contactless tunnel-inspection unit with 1.12 m coverage.[6] No CAD model was available from the manufacturer, and a direct request went unanswered, so the geometry was rebuilt from published dimensions and product imagery with internal volumes left hollow, so the modeled mass matches the published 35 kg rather than a solid-body estimate.[7]

The final RobotStudio flange-load record is 70.2 kg with center of gravity at (−3.1, 26.6, 412.1) mm, principal inertias of 7.611, 5.620, and 4.600 kg·m², and AOM quaternion [0.9926, 0.1209, 0.0048, 0.0073]. Those values pair the CAD-measured tool-side assembly and adapter with the vendor-specified 6.44 kg QC-160 Master mass so imported solid-model mass errors do not enter the robot load definition.

Three load cases, one governing result.

The platform was screened in Fusion using ABB's published foundation loads applied at the robot mounting plate, in both endurance and emergency-stop cases, across the two worst reach directions.[1][2]

Supports are idealized at the bogie adapter plates using a pin-and-roller-style scheme: the front adapter is fixed, and the rear is held vertically and laterally but free to move longitudinally. Fixing both ends would have made the frame artificially stiff. All analyzed solids use ASTM A572 properties with a 345 MPa yield; ASTM A500 Grade C remains the intended HSS specification for a future detailed design.

The first version of this study contained an error worth documenting. Reviewing the exported report showed that three moment vectors had resolved onto the wrong axes when they were defined by picking reference edges, which meant the predicted worst case, the side-tipping emergency stop, had never actually been solved. Correcting the axes and re-running increased the governing stress from 47.6 MPa to 85.1 MPa. The design still passed, but the first result was not the result it appeared to be.

Stability was the harder problem.

Anti-tip stability was calculated separately as a rigid-body moment balance about the rail contact line at half gauge, comparing resisting moment against overturning moment. Before the deck was added, the realistic-bogie emergency-stop sensitivity case fell below the 1.5 target and indicated approximately 8.2 kN of hold-down.

That result conflicted with an operating requirement established through a subject-matter-expert interview with an MTA conductor: the inspection vehicle must remain in motion while scanning, so stop-and-clamp operation is not viable.[18] The concept response is to define intended SafeMove speed, acceleration, and reach limits rather than clamps; application-specific control validation remains open. Separately, adding the measured 1,875 kg deck to the rigid-body balance moved every screened case above the project threshold with no hold-down required.

The 1.515 result clears the project criterion by about one percent. It is reported as a concept-level screening pass, not a certification, and it applies all published load maxima simultaneously even though the manufacturer states those maxima do not normally occur together. The workbook still carries the earlier 68 kg tool input and a conservative 2,500 mm lateral-reach placeholder; reconciliation to the final 70.2 kg load record and a measured worst RobotStudio pose remains open.

The inspection cycle, running.

The final RobotStudio station carries the tunnel, rails, platform, ABB IRB 6750S-215/3.9, OmniCore V250XT controller, tool changer, tooling, and tool stand, with the final 70.2 kg payload, center of gravity, inertia, and AOM active in the tool definition.

The inspection path sweeps fifteen targets from the platform-level sidewall to the tunnel crown at a constant 75 mm standoff, holding the radar face normal to the lining, at 200 mm/s with blended motion so the sweep is continuous rather than stopping at each target. Lower-wall and floor coverage are not claimed. Standoff was chosen within the sensor's published contactless capability.[6]

CollisionSet_1 checked the complete GPR tool against the tunnel geometry, the platform assembly, the robot links and base, and the V250XT cabinet. The logged program completed with zero collision events, zero singularity warnings, and no joint-limit or motion-supervision events across repeated cycles, and the sweep was also reviewed visually for contact. The train reference, wayside equipment, and cable routing were not modeled and are therefore outside that check. A signal-driven attach and detach sequence also demonstrates conceptual docking and drop-off logic, but a small reattachment offset remains; autonomous re-pick is not claimed as validated.

What was tried, what failed, and what changed because of it.

This is the design trail rather than a highlight reel. Several of these entries are corrections found late, and they are included because catching them is part of the engineering.

Ten key models from the assembly, open to inspect.

These are live Fusion 360 viewers. Drag to orbit, scroll to zoom, and use the browser panel to explore components.

Sources behind the design decisions.

Vendor documentation, transit references, published research, and public models used to develop and check this project. The full research record, including prior art and the assumptions log, is maintained separately.