Fixing Bus ETAs: AI Lessons for Supply Chains

On Feb 7, 2026, Singapore’s Land Transport Authority (LTA) said restoration of the bus Expected Time of Arrival (ETA) system was more than 90% complete after a month of glitches that began on Jan 10. The fix wasn’t a quick “turn it off and on again” story. It involved clearing memory cache, manually updating firmware, and in some cases replacing transmitters across roughly 4,000 buses.

If you work in logistics, fleet operations, retail, or any business that depends on real-time status updates, this incident should feel uncomfortably familiar. When live tracking breaks—whether it’s buses, delivery vans, cold-chain sensors, or warehouse robots—customers don’t just lose information. They lose trust.

This post is part of our “AI dalam Logistik dan Rantaian Bekalan” series, where we look at how AI improves routing, warehouse automation, demand forecasting, and end-to-end supply chain performance. The LTA bus ETA recovery is a clean case study in a single, blunt lesson: AI is only as good as the data pipeline feeding it—and resilience is a design choice, not a nice-to-have.

(Source story: https://www.channelnewsasia.com/singapore/lta-bus-eta-arrival-timing-system-more-90-cent-restored-5909491)

What actually went wrong—and why it matters to businesses

The direct issue, according to LTA, was technical problems that prevented buses from transmitting location data to a central server. Engineers and the system contractor identified a “memory cache build-up” on on-board systems, affecting about half the fleet at one point.

Here’s the business translation: the “AI-looking” front end—bus stop displays and transport apps—wasn’t the real problem. The weak link was the edge-to-cloud data path:

• Edge device on each bus (transmitter / on-board unit)
• Firmware + memory management
• Connectivity to central servers
• Aggregation and estimation logic (ETA)
• User-facing apps and displays

When that chain breaks, the customer experience collapses fast. In supply chain terms, it’s the same pattern as:

• Delivery ETAs freezing because telematics devices stop sending GPS pings
• Warehouse dashboards going stale because scanners desync after an update
• Cold-chain compliance alerts failing because sensors fill local storage

My take: many companies over-invest in dashboards and under-invest in the boring plumbing—device health, telemetry reliability, and recovery procedures. The plumbing is what customers feel.

The “longer headways” problem is a perception problem

LTA noted commuters saw missing arrival timings or long wait times displayed. Even if buses kept running “at their usual frequencies,” perception drives behaviour:

• People leave earlier “just in case”
• They switch modes (ride-hail, MRT, walking)
• They complain loudly because the system promised certainty and delivered confusion

In logistics, that’s the difference between:

• “Your parcel is coming today” vs “Your parcel is coming between 9am–10pm”

The second one might be operationally true, but it’s commercially damaging.

The restoration playbook: what we can copy in fleet and logistics systems

LTA’s remediation steps were concrete: clear cache, manual firmware updates, replace transmitters where needed. It’s a reminder that reliability is not a single fix—it’s a set of operational disciplines.

Below is a practical recovery-and-prevention playbook you can adapt for logistics fleets, warehouse automation, and supply chain visibility platforms.

1) Treat edge devices as production systems (because they are)

The incident required technicians to physically service devices. That’s normal in any distributed fleet, but companies often act surprised when “IoT” needs hands-on ops.

Do this in your environment:

• Maintain an asset registry: device model, firmware version, last check-in, last update, operator/vehicle mapping
• Define a device health score (battery, storage, temperature, connectivity uptime)
• Set alerts for silent failures (no heartbeat) and degraded failures (heartbeats without useful payload)

Snippet-worthy rule: If you can’t measure device health, you’re not doing real-time tracking—you’re doing wishful thinking.

2) Build rollback and staged rollout into every firmware update

LTA’s issue involved firmware updates and cache behaviour. In logistics, OTA updates can quietly introduce instability if you don’t stage releases.

A robust rollout pattern:

1. Pilot on 1–2% of vehicles/devices
2. Expand to 10% after 48–72 hours of stability
3. Push to 50% only when you’ve observed peak-hour conditions
4. Complete rollout + keep one-click rollback ready

If you manage warehouse systems, apply the same discipline to WMS plugins, scanner firmware, conveyor PLC settings, or robot software.

3) Engineer “graceful degradation” instead of total failure

When ETAs disappear, users feel abandoned. But you can often provide a fallback estimate even when live signals are missing.

Examples that work well:

• Use last-known position + route history to provide a confidence-banded ETA (“~6–10 min, lower confidence”)
• When GPS pings drop, use scheduled headway + historical variance to avoid blank screens
• Provide a plain-language status: “Live tracking temporarily unavailable; using timetable estimate”

For supply chain visibility:

• If live telematics fails, estimate via scan events, depot departure times, and historical transit time distributions
• If warehouse sensors fail, switch to process-based inference (“picked but not packed yet”)

The goal isn’t perfection. It’s continuity.

Where AI fits: from “predicting ETAs” to preventing the next outage

Most teams talk about AI as “predictive ETAs.” That’s only half the story. The higher ROI is often in predicting system failure and improving resilience.

Predict failures before users notice (AIOps for fleets)

LTA described cache build-up affecting performance. That’s a classic anomaly pattern.

AI can help by learning leading indicators of device trouble:

• Rising memory consumption over time
• Increased packet loss or retransmits
• GPS jitter spikes (device struggling)
• Longer intervals between successful uploads

You don’t need sci-fi AI here. A practical stack might include:

• Simple anomaly detection (e.g., seasonal decomposition + thresholds)
• Classification models to predict “will fail within 7 days”
• Automated ticket creation + route technician visits efficiently

This is very aligned with AI dalam logistik dan rantaian bekalan: fewer breakdowns means better on-time performance, fewer exception-handling costs, and more reliable customer promises.

Better routing and dispatch decisions depend on trustworthy real-time data

Real-time data drives real-time decisions:

• dynamic route optimization
• fleet rebalancing
• exception management (late deliveries, missed pickups)
• warehouse wave planning

When the data stream becomes unreliable, AI can amplify the problem—optimizing based on wrong inputs.

Hard stance: if your data quality isn’t monitored like uptime, stop automating decisions at scale. Automate after you can trust your telemetry.

A system upgrade plan is a signal, not a footnote

LTA said it has started upgrading the fare and bus fleet management systems, to be completed within two years. That’s the right move: don’t just patch; modernize the stack.

In business terms, this is the “platform renewal” decision:

• replace brittle integrations
• update device lifecycle management
• standardize telemetry formats
• implement better observability and incident response

If you run logistics operations in Singapore, this is a helpful benchmark: public infrastructure teams are treating transport data as a long-term product, not a side feature.

A practical checklist: making real-time supply chain tracking resilient

If you own operations, product, or IT for logistics/fleet/warehouse systems, use this checklist in your next monthly review.

Data reliability and observability

• Define SLOs for tracking availability (e.g., “live location available for 95% of fleet during service hours”)
• Track missing data rate and stale data rate separately
• Implement end-to-end tracing (device → gateway → server → app)

Incident response muscle

• Maintain runbooks for top 10 failures (connectivity outage, firmware bug, server overload, cache build-up)
• Run quarterly game days: simulate a telemetry drop and test recovery time
• Ensure you can do targeted fixes (update only affected devices, not the whole fleet)

AI governance (yes, even for ETAs)

• Version your ETA models and document feature dependencies
• Monitor model drift when operational patterns change (new routes, new hubs, festive season peaks)
• Use confidence scores in customer-facing ETAs to avoid false certainty

February in Singapore tends to bring busy commuting patterns after year-start ramp-ups, plus operational variability around school and work schedules. That’s exactly when brittle systems get exposed—because demand is normal-high and tolerance for uncertainty is low.

What commuters can teach your customers (and your sales team)

Commuters didn’t just want an arrival time. They wanted to plan their lives. Your customers want the same:

• A retailer wants to plan staffing for receiving
• A clinic wants to schedule deliveries of temperature-sensitive supplies
• A contractor wants materials to arrive before workers are on-site

So the promise isn’t “we have tracking.” The promise is:

Reliable ETAs reduce wasted time, reduce support tickets, and make operations calmer.

When systems fail, proactive communication matters almost as much as the fix. LTA’s public updates (progress percentages, causes, steps taken) are a useful template for business incident comms: state what’s happening, what’s impacted, what you’re doing, and when the next update will come.

What to do next if you’re building AI for logistics in Singapore

If you’re investing in AI business tools for logistics and supply chain—route optimization, warehouse automation, or real-time visibility—take this bus ETA incident as a design prompt:

1. Audit your data pipeline from edge devices to dashboards
2. Add device health analytics before adding more predictive models
3. Design fallback estimates so customer experience doesn’t drop to zero
4. Put staged rollouts and rollback into your update process

The bigger question worth sitting with: when your tracking system is stressed, does it fail loudly and clearly with a safe fallback—or does it fail silently and confuse everyone?