Spot at the rail factory

A Boston Dynamics Spot, custom-rigged with a Slalom sensor stack, walking the floor of a 24/7 rail factory and feeding defect detections into the work-order flow. The robot’s job is to see what humans can’t keep up with, in an environment that moves faster than the map.

Context

Hitachi runs a rail-manufacturing facility built on Industry 4.0 principles: modular, instrumented, reconfigured every shift to match the next production run. Inspection at that scale is a navigation problem before it’s a sensor problem. The shop floor reorganizes itself between shifts. Carts, jigs, scaffolds, half-built railcars all move. The only fixed reference points in the room are the structural columns holding up the roof.

Hitachi wanted an autonomous inspection capability that could keep up. Slalom led the program end to end, concept through live deployment. The work landed in three connected stretches across about a year.

1. Santa Clara, R&D campus

We started with a co-development sprint with Hitachi Digital Services at Hitachi’s R&D campus in Santa Clara. Two goals: align on the operational problem we were actually solving, and prototype the technical pattern fast enough to show on a railcar.

The platform: a Boston Dynamics Spot running custom Slalom software on Spot’s onboard Core I/O. The connectivity sketch: a private 5G network at the production facility (with Ericsson) bridging robot-side compute to Hitachi Digital Services in the cloud. The backend sketch: a real-time digital twin that turns detections into work orders. We built the partial-scale railcar mock-up on the R&D floor and used it to debug the robot’s first Autowalk routes, the sensor stack’s first defect captures, and the data pipeline end to end before any of it touched the production line.

2. Slalom NY lab, integration

We took the platform back to the New York lab to harden it. Several sprints. Camera calibration first, because nothing downstream works if the cameras don’t agree on what they’re seeing.

We tuned the sensor stack for railcar interiors, taught the navigation stack to handle a floor that wouldn’t sit still, and trained CV defect detection on real Hitachi rail components. We also shipped a human-in-the-loop review surface where operators sign off on detections before they trigger work orders. The review surface is the part most teams underestimate. A robot that flags defects nobody trusts is a robot that doesn’t ship.

3. On-site, dock and first walk

Three to four months on the Hitachi factory floor. We got the platform off the truck and into the bay it would live in for the rest of the program.

The first Autowalk teach pass is a slow, pedantic affair. Drive the robot through the route with a tablet, drop a waypoint, take a reading, drive to the next, drop another. Then play it back and watch what breaks.

The factory was the adversary. Carts moved. Scaffolds moved. AGVs threaded routes the robot hadn’t seen yesterday. Every shift change put the map in question, and the robot had to know what was a fixed feature (a column) versus what was a guest (a jig).

5. Inspection, the actual job

Once the robot could move, it had to look. The mission ran the same way every time: walk a pre-programmed route, hit each inspection station, point the PTZ payload at the target, capture high-resolution stills of the asset surface, push them to the edge server over private 5G for CV analysis, log the result against the asset record, and continue.

The inspection targets were prosaic and merciless. Rivet lines on side panels. Weld continuity along seams. Brake-caliper alignment on wheelsets. The kind of finish-quality call that a human inspector spots in a glance and that a CV model has to learn from a thousand examples of.

The data pipeline did the heavy lifting. PTZ stills are big. Hundreds per mission, multi-megapixel. The private 5G network shipped them off the robot to a local edge server, where the CV models ran inference, wrote a per-asset condition record, and updated the digital twin. From there, anything flagged surfaced in the human-in-the-loop review queue. Operators verified, approved or rejected, and approved detections dropped into Hitachi’s work-order system for the quality team to action.

6. Production

The robot ran 24/7 alongside humans and existing AGVs, executing routes triggered from a tablet controller and feeding the digital twin in real time.

Outcome

Live deployment in Hitachi’s rail-manufacturing facility, 24/7 alongside humans and AGVs
End-to-end Slalom integration on the Spot platform: sensor stack, navigation, CV defect detection, HITL review, work-order tie-in
Private 5G bridges robot-side compute to a local edge server; edge server runs CV inference and populates the digital twin and the work-order management system
Continuous review surface where operators verify detections and trigger work orders for the quality team
Presented at India Mobile Congress and Mobile World Congress as a flagship Industry 4.0 deployment