The Railira platform is a plant-edge inference runtime on NVIDIA Jetson, a plant orchestrator with a policy engine, specialist agents for every step of the rolling-stock plant, and an as-built train-and-plant twin that rehearses each shift. Everything an agent does is typed, bounded and logged.
Illustrative scenario: Car 07 of a 4-car metro consist · shell line B · bogie shop · booth 2 · marriage station, one car, one approval gate
Perceive the panel, body and car, plan the shift, act inside limits, inspect continuously, optimise quality and takt, log everything, retrain. The layers below are the design, not a description of a shipped system.
Rail PLM and MES, welding and coating robot controllers, laser trackers and scanners, torque tools, wiring testers, brake-test rigs and commissioning systems, coating-booth instruments, AMR fleets and standards documents. Signals are normalised into one time-aligned car and work-package record.
DeepStream and Holoscan process weld, coating, metrology, torque, thermal and robot streams on Jetson Orin. TensorRT-optimised models detect weld and coating defects and geometry risk; time-series models predict torque escapes, takt loss and cell health. Triton serves them with versioned rollback.
The plant orchestrator assigns work to specialist agents. Agents retrieve control plans, work instructions and standards through permission-aware retrieval, evaluate action envelopes, ask for approval when required, and write back to robots, MES, torque tools or work packages only when the policy layer allows.
The Omniverse twin receives as-built shell geometry, running-gear data and coating state, compares them with as-designed intent, and tests candidate weld sequences, geometry strategies, marriage plans and build orders before the next shift. GPU FEA for weld distortion and geometry; point-cloud metrology.
Training on DGX or HGX, fleet model registry, versioning and network reporting. Federated, IP-isolated learning across plants is planned and waits on customer data-rights agreements. Per-tenant isolation throughout.
A review console for body, coating and quality engineers to approve, correct and give feedback. Corrections train the models. An append-only, assurance-grade IRIS and quality audit log records every agent action from day one.
The core user flow, in order. Outputs are reviewed by a carbody, coating or quality engineer whose corrections train the models.
Every Railira run is an ordered, inspectable sequence. This is run_9b4e27: Car 07 of a 4-car metro consist · shell line B · bogie shop · booth 2 · marriage station. Scenario completes in 10 steps with one approval gate. Shell geometry held inside tolerance after a twin-ranked weld sequence, one heat-input change approved by the shift weld engineer, two torque traces routed to rework before brake test, the livery cure window protected in the sequence, marriage interface inside tolerance, car released with full genealogy.
Plant Orchestrator · Pulled the work packages for car 07 from MES and PLM: aluminium extrusion shell, EN 15085 weld class per joint, control-plan geometry tolerances, bogie torque specifications, livery recipe and cure window, fit-out and marriage sequence. Locked the approved control-plan envelope for the shift.
Line-and-Logistics · Sequenced the shift against live readiness: booth 2 cure window, bogie B-14 readiness, brake-test slot at 13:40, AMR fleet state and kit inventory. The plan moves car 07 marriage behind car 06 to keep the cure window and the test slot.
Train Twin · Simulated 18 candidate weld sequences for the roof-to-sidewall joints on the as-built twin and ranked them on predicted distortion, door-aperture geometry and heat input. Sequence #6 keeps the aperture inside tolerance with 2% less heat input than the standing sequence.
Weld-and-Structure · Recommended sequence #6 and the matching clamping plan to the shell line. Seam vision and laser-tracker geometry confirmed the roof-to-sidewall joints inside the control-plan tolerance on the first four seams.
Plant Orchestrator · Sequence #6 lowers heat input on the roof joints 2% below the standing weld procedure. Heat-input changes stay with a named engineer, so the step was held for the shift weld engineer and approved at 06:12:44 with the twin evidence attached.
Bogie-and-Running-Gear · Press-fit curves and torque traces on bogie B-14 read at the station. Two traction-motor mounting traces fell outside the expected profile and were routed to rework before brake test; the remaining safety-critical joints verified to spec.
Coat-and-Livery · Booth 2 humidity and temperature, film thickness and flash time watched through the livery pass. One masking misalignment on the door band flagged by booth vision before the second coat; the cure window held and handed to the sequence as a hard constraint.
Marry-and-Couple · Carbody and bogie scans compared against target interface geometry before the lift. Fit-out readiness confirmed by Fitout-and-Interior for doors and glazing. Marriage guided inside tolerance; coupler and gangway checks verified and written to the twin.
Quality-and-Anomaly · Weld and coating vision, in-line metrology, torque, wiring continuity and brake-test results fused at the edge across the car. Every finding localised to station, work package and shift, with the two torque reworks and the masking flag already closed.
Conformance · Car 07 released: genealogy written (weld records per joint, geometry scans, torque traces, coating recipe and cure record, fit-out checks, marriage interface, approvals, test results) and dispositioned against the control plan and IRIS / ISO/TS 22163 criteria with cited evidence for the quality engineer's sign-off.
Every tool invocation, argument and result is written to an append-only, human-readable log, and every reasoning step is expandable. Nothing about a run is hidden from the plant.
Plan, thought, action, observation, exactly as the orchestrator would record it.
Build car 07 through shell welding, bogie assembly, livery, fit-out and marriage on the shift plan, inside the control plan, with every safety-critical joint to spec and the cure window protected.
Sequence the shift first, so the cure window and the brake-test slot are protected. Simulate the roof-to-sidewall weld sequence on the twin before the first seam. Read every torque trace at the station. Fuse inspection continuously. Hold any heat-input change for a human.
The standing weld procedure runs the roof joints hot enough to guarantee penetration on a cold fixture. With the clamping plan from the twin the same penetration holds at 2% less heat input and the door aperture stays inside tolerance, because the distortion is sequenced out rather than reworked out.
twin.simulate(candidates=18) returns sequence #6: roof joints re-ordered, clamping plan 6, heat input down 2% on the roof joints.
Sequence #6 predicts door-aperture deviation 0.6 mm against a 1.5 mm tolerance. Three candidates scored lower on heat input but pushed predicted sidewall distortion past the limit and were discarded.
Recommend sequence #6 to the shell line; six recommendations permitted; the heat-input change routed to the shift weld engineer with the twin evidence.
Two traction-motor mounting traces on bogie B-14 fell outside the expected profile at 07:40:12. Routed to rework before brake test; the slot held. One masking misalignment on the door band flagged in booth 2 before the second coat.
Scenario complete. Shell inside tolerance, bogie joints to spec after rework, livery inside the cure window, marriage interface inside tolerance. Car 07 released with full genealogy.
Two agents will want the same sequence slot. The orchestrator arbitrates on the run goal, not on who asked first, and the handoff is logged like any other step.
Line-and-Logistics wants car 07 on the marriage station at 10:40 to hold the 13:40 brake-test slot.
Coat-and-Livery wants car 07 held until 11:05 because the second coat's cure window does not close until then.
A broken cure window is a repaint; a 25-minute slip still reaches the test slot. Under the run goal "livery inside the cure window", Coat-and-Livery wins. Marriage moves to 11:10, car 06 takes the earlier station time.
Sequence re-solved with the cure window as a hard constraint. Brake-test slot at 13:40 held; AMR routes updated.
Illustrative scenario.
A rolling-stock plant does not go from manual to unattended in one step. Railira makes the level explicit, auditable and reversible at any time. The first release runs shadow mode, then assist mode, then bounded autonomy on low-risk actions. Weld parameters, schedule release and car release stay with named humans until acceptance is earned.
| Level | What the agent does | What the human does | Where it sits in a pilot |
|---|---|---|---|
| L1 · Shadow | Reads weld, metrology, torque, coating, wiring, test, robot and MES signals; predicts and recommends; writes nothing | Runs the line as today; compares predictions to inspection and engineer outcomes | Shadow mode; baseline measured from the plant's own records |
| L2 · Assist | Proposes weld-sequence, geometry, coating, torque and sequencing moves with evidence, confidence, envelope and rollback path | Approves each move in the review console | Assist mode; accuracy and ROI proven against the baseline |
| L3 · Bounded | Executes pre-approved low-risk actions: inspection routing, rework priority, kit feeding, robot path variants and parameter adjustments inside approved bounds | Approves heat-input and weld-sequence changes, schedule release and car release | Graduated autonomy, after audited accuracy and ROI and safety gates |
| L4 · Scaled | Supervised weld-parameter adaptation inside pre-approved control-plan envelopes, across lines and plants | Sets the envelope; reviews the shift record and the evaluation results | Planned; after IRIS, plant and OEM acceptance is earned |
A seam is welded in seconds and a booth cannot wait for the cloud. Perception has to be local, deterministic and fast, so Railira runs GPU inference at the plant edge and keeps training, simulation and optimisation in the cloud or on-prem.
8 to 32 cameras per process area plus metrology, torque, thermal, robot and booth-instrument streams fused at the plant edge. Sub-50 ms weld and seam alerts and sub-150 ms coating and metrology prompts are the design targets. Jetson Thor is planned for higher sensor density.
Weld-defect, geometry-risk, coating-risk, point-cloud and takt-risk models served at the plant edge with per-model versioning and rollback, routed by latency, tenant, sensitivity and cost.
Weld and coating defect vision, metrology classification, control-plan reasoning and FEA-surrogate models fine-tuned on plant data and engineer corrections. 100 to 500 GPU-hours per plant during onboarding is the design target; dedicated capacity is planned.
As-built carbody, coating, marriage and line-flow twin. Weld distortion, geometry build-up, robot reach, coating coverage and layout simulated before the shift. GPU FEA and point-cloud metrology are the workloads; OVX scale is planned.
Rare weld defects, coating contamination, geometry deviations, occluded seams and torque-escape cases generated and validated against real inspection distributions before promotion. Planned.
Line sequencing, JIS feeding, AMR and AGV routing, buffer management and commissioning-slot scheduling against live readiness.
Robotic weld, coating and assembly cells and AMR or AGV flow; path variants validated in simulation before release to the floor.
Telemetry features from weld parameters, metrology, coating readings, torque, non-conformance and takt history for retraining, drift detection and network analytics.
Railira reads and writes through the systems already on the floor. No rip and replace, no parallel historian, no new HMI to learn. Vendor names are examples of systems found in rolling-stock plants, not certified integrations; connector availability is confirmed per site.
Rail PLM, MES and sequencing suites (Siemens Teamcenter, Opcenter and Tecnomatix, Dassault DELMIA, Rockwell, iBASEt, Critical Manufacturing class systems)
Work packages, control plans, sequence reads; genealogy and schedule write-back
Robot controllers and process controllers (FANUC, KUKA, ABB, Yaskawa, Comau, Dürr, Kawasaki class systems)
Parameter and path reads; bounded parameter and sequence recommendations
Laser trackers, point-cloud scanners and vision systems (Hexagon, Zeiss, FARO, Cognex, ISRA Vision, Perceptron class systems)
Geometry scans and inspection results per car
Torque tools and controllers (Atlas Copco, Bosch Rexroth, Stanley class systems)
Torque traces per joint, rework routing
Wiring testers, brake-test rigs, traction-test and ETCS, ERTMS or PTC commissioning systems
Continuity, brake and signaling results per car
Booth instruments, thickness gauges, environmental sensors, recipe systems
Film build, flash, cure, humidity and temperature
Plant-simulation and twin tools (Siemens Tecnomatix Plant Simulation, Dassault DELMIA, Visual Components)
Model and result exchange with the train-and-plant twin
Fleet managers and dispatch APIs
Fleet state, dispatch recommendations
Railira writes to production systems. Every capability is scoped, every write is policy-checked against the approved control plan, and every action is written to an append-only audit log the plant owns.
| Item | Scope | Status |
|---|---|---|
| SOC 2 Type I and OEM cyber review | Cloud control plane | QUEUED Planned, first 6 months |
| SOC 2 Type II | Cloud control plane | QUEUED Planned, 6 to 12 months |
| SSO and RBAC | Console and approvals | QUEUED Planned, 6 to 12 months |
| IRIS and regulatory assurance packages | Disposition and genealogy evidence | QUEUED Planned, 6 to 12 months; not a certification claim |
| Standards-cited reasoning | Control plans, EN 15085, IRIS / ISO/TS 22163, TSI, AAR, APTA, torque and brake-test procedures | RUNNING In the design, every answer cites its source |
| Per-tenant and program isolation | Designs, control plans, models, memory | RUNNING In the design from day one |
| On-prem and plant-edge deployment | Regulated plants | RUNNING Design option |
The questions plant directors, body and coating engineers and quality leads ask in the first meeting.
Only within an explicit tool allow-list with per-tool rate and magnitude limits, inside the approved control-plan envelope, and only at the autonomy level your plant has set. Design partners start in shadow mode, where Railira writes nothing. Heat-input and weld-sequence changes, schedule release and car release stay with a named approver.
Control returns to the robot, PLC or station's last known-good state. Railira is a supervisory layer on top of the controllers, safety systems and interlocks you already run, never a replacement for them, so an outage degrades the line to its current way of running, not to a stop.
Recommendations start from the control plan, the work instructions and the twin in the first weeks. Weld, geometry, coating and torque prediction improve as line history and labelled outcomes accumulate per car type. Onboarding is sized at 100 to 500 GPU-hours per plant and 50,000 to 250,000 synthetic images or scan variants per program, all design targets.
Only if you choose cloud training. Designs, control plans, program data, process models and quality records are isolated per tenant and never used to train another customer's models. Fully on-prem deployment at the plant edge is available for regulated plants, and cross-plant learning is planned and waits on data-rights agreements.
You are, the same as with any control strategy, which is why every write is policy-checked, bounded, logged and reversible, and why anything above your risk threshold waits for a named approver. The audit log records the request, the evidence, the envelope applied and the human decision.
A paid pilot on one process or line area, scoped to a single wedge with a success metric written into the order form: weld or coating rework reduction, geometry escapes, torque escapes, line stoppage minutes or first-time-through. It starts in shadow mode against a baseline from your own records, moves to assist mode when the engineer is satisfied, and expands only after the number moves.
Start with one carbody, bogie, coating, fit-out or sequencing process and one measurable baseline. A paid pilot on weld and geometry, torque assurance, livery defects or line sequencing shows the number before you commit further.
↑↓ navigate↵ openesc close
