Field Log · Selected Work
Case notes from the floor — anonymised, measured, not promised.
These engagements are described without client names so we can share what we learned on site. Metrics are past results from specific environments. They are not guarantees for your fleet, cell or pilot. Smart Robo AI is a robotics-autonomy engineering studio in Vancouver — not a course, wellness brand, or financial robo-advisor.
Canadian manufacturer · AMR fleet · Fleet analytics
When aggregate uptime hid a shift problem
A Vancouver-area manufacturer operated twelve autonomous mobile robots across day and night shifts in a 40,000-square-foot assembly annex. Executive dashboards reported ninety-one percent fleet uptime — acceptable against their internal SLA. Our autonomy review plotted telemetry per shift and found night-shift localization failures clustered around a staging lane reconfigured six weeks earlier. Sensor fusion weights favoured day-shift lighting profiles; SLAM drift spiked when overhead arrays dimmed for energy savings after 22:00.
We ran a sim-to-real validation pass in a digital twin seeded with night-shift lidar captures, retuned fusion parameters with explicit confidence bands, and deployed a guarded policy update with rollback hooks. Night-shift incident rate dropped in a four-week measurement window. Day-shift performance held steady. The client retained layout-change notification as an operational control. We did not guarantee throughput gains — maintenance discipline and future layout edits remain variables outside our scope.
National logistics operator · Grasp policy · RL control
Simulation success, cardboard reality
A national logistics operator needed a grasp policy for mixed-SKU totes on a pick-and-place line feeding an AMR loop. Reinforcement learning in simulation reached high success rates on rigid corrugated boxes with clean edges. Floor trials on crushed cardboard, label overlap and inconsistent tape seams told a different story — success fell below the operator's internal threshold within the first shift.
We rebuilt the neural perception stack with floor-captured training data, ran edge-case discovery sessions with shift leads, and added human-robot interaction guardrails at the handoff zone where operators intervene on failed grasps. The revised policy shipped with documented failure modes, confidence ranges per SKU class and a safety briefing artefact for the client's joint health-and-safety committee. Simulation metrics alone would have misled a production sign-off.
Regional distributor · Warehouse AMR · Predictive maintenance
Bearing wear flagged before the aisle stop
A regional distributor running eight warehouse AMRs in a BC interior facility saw recurring drive-unit failures that appeared random in their CMMS. Aggregate vibration data looked flat. We ingested motor current signatures, wheel odometry residuals and incident timestamps — then aligned them to route segments and ambient temperature logs from the loading dock.
Predictive maintenance models flagged elevated bearing wear probability on two units servicing the cold-storage transition corridor. Recommended maintenance windows were scheduled during planned downtime. Subsequent failures on those units did not recur over a twelve-week follow-up. Other units continued on standard intervals. The engagement did not eliminate all mechanical risk — it made one failure mode visible earlier than reactive repair.
Film-stage automation shop · HRI · Safety assurance
Shared workspace where cues mattered more than speed
A False Creek-adjacent film-stage automation shop integrated a collaborative arm for prop handling between human craftspeople and a mobile base. The OEM's default motion planner prioritised cycle time; near-miss reports increased when lighting changed between rehearsal and shoot configurations. We conducted a human-robot interaction and safety assurance review mapped to their zone layout — not a full facility certification.
Deliverables included revised approach speeds in shared zones, audible and visual cue timing aligned with stage call patterns, and a test protocol for lighting-change rehearsals. Incident reports dropped in the client's internal log over an eight-week production block. We documented residual risk explicitly: stunt choreography and unscripted personnel movement remain human-supervised responsibilities.
Want a structured read on your stack? An autonomy review is the usual entry point — scoped timeline and CAD bands follow discovery.
Your floor has its own edge cases.
Share your fleet, cell or pilot context — we will respond with an honest scope read and what simulation plus floor data can support.
Request an autonomy review