Grab’s Robot Delivery Bet: What Singapore Can Copy

Grab’s acquisition of Infermove (announced 7 Jan 2026) isn’t a flashy “robots everywhere” headline. It’s a practical signal: Singapore-based companies are treating AI dalam logistik dan rantaian bekalan as an operations upgrade, not an innovation side-project.

Last-mile delivery is where logistics economics get painful. Costs pile up in small, repeated actions: finding the right block, navigating tight corridors, waiting for lifts, handling hand-offs, and dealing with unpredictable pedestrian traffic. Humans are great at improvising—but improvisation doesn’t scale cleanly.

What I like about this move is that it’s not just about autonomy. It’s about data, training loops, and operational design—the stuff that actually determines whether AI in logistics delivers ROI.

Why Grab bought Infermove (and why it matters)

Grab acquired Infermove, a Chinese AI robotics company, to strengthen automated delivery across first-mile and last-mile logistics. Infermove will operate independently, with founder Aaron Lu reporting to Grab’s CTO, Suthen Thomas.

Here’s the clear takeaway: Grab is buying capability, not just hardware. Sidewalk robots are visible, but the real asset is the AI stack that learns from messy real-world conditions.

From the source article, Infermove focuses on:

• Autonomous driving systems for unstructured environments (the opposite of neat factory floors)
• Mobile manipulation robots (robots that don’t just move, but can handle objects)
• Sidewalk delivery robots with upper-limb manipulation
• A proprietary Rider Shadow System to collect training data
• Training via imitation learning and reinforcement learning

That combination points to a specific strategy: build robots that can handle the chaos of the last mile, then improve them continuously using real-world data.

The real problem in last-mile logistics: variability, not distance

Last-mile delivery isn’t “the last 3 km.” It’s the last 30 decisions.

In Singapore (and most dense cities), delivery complexity comes from:

• Mixed environments (sidewalks, shared paths, drop-off lobbies)
• Pedestrians, PMDs, cyclists, construction zones
• Lift etiquette and building access rules
• Weather disruptions and peak-hour crowding
• Address ambiguity (“behind the guardhouse”, “left of the loading bay”)

AI dalam rantaian bekalan works best when the bottleneck is decision-making at scale. Last-mile is full of micro-decisions. If you can standardise and automate even a slice of them, you reduce:

• Cost per delivery
• Missed delivery rates
• Rider idle time
• Variance (which is what breaks planning)

Robots won’t replace all riders anytime soon. But they can take on specific routes, time windows, or building types where the economics are predictable.

Why “unstructured environments” is the key phrase

Plenty of autonomy systems work in controlled areas (warehouses, campuses). Unstructured environments means the robot can’t rely on perfect lanes, perfect maps, or predictable obstacles.

That’s why Infermove’s approach matters: training on real-world mobility data (including delivery riders’ electric scooters) is a direct attempt to solve the hardest part—generalisation.

Infermove’s data advantage: the “Rider Shadow System” approach

Infermove collects training data using last-mile mobility devices and then trains robots through imitation learning and reinforcement learning.

This is a smarter route than building a robot and hoping it learns later.

Imitation learning vs reinforcement learning (simple, useful view)

• Imitation learning: the robot learns by copying expert behaviour (for example, how a rider slows near a blind corner).
• Reinforcement learning: the robot learns by trial-and-error within constraints, improving policy over time (for example, choosing safer paths that also minimise delays).

A good production setup uses both:

1. Learn baseline behaviour fast via imitation.
2. Improve edge cases via reinforcement.

In last-mile delivery, edge cases are everything. A robot that handles 95% of situations but fails in the remaining 5% can still be operationally expensive if that 5% causes rescues, refunds, and reputational damage.

What Singapore businesses should copy from this

Even if you’re not building robots, the method applies to many logistics processes:

• Capture “shadow data” from the way your best staff operate.
• Turn it into training signals.
• Deploy automation that improves with feedback.

This is how you move from one-off pilots to compounding operational gains.

Where robotics fits in AI supply chain strategy (beyond PR)

Robotics becomes valuable in supply chain automation when it connects to planning and control systems. A robot is not a standalone product; it’s a node in your logistics network.

For Grab, integrating robotics can strengthen:

• Dispatch logic: matching jobs to humans vs robots based on probability of success
• Route optimisation: learning which paths actually work on the ground
• Service level reliability: fewer delays due to rider shortages or peak demand
• Safety and compliance: more consistent behaviour under defined rules

For other Singapore companies—retailers, 3PLs, manufacturers—robotics might show up differently:

• Yard movement and facility-to-facility transfers
• Campus delivery inside business parks
• Assisted picking and internal replenishment
• Micro-fulfilment for dense neighbourhoods

A stance I’ll take: most companies should not start with robots. They should start by making their workflows “robot-ready”: clean data, clear SOPs, measurable handoffs, and exception handling.

A practical adoption map for AI in logistics (Singapore context)

If you’re leading ops, you need a path from “interesting tech” to “operational KPI.” Here’s a sequence that works.

Step 1: Standardise the delivery/fulfilment truth

If your delivery timestamps, status codes, or location fields are inconsistent, AI will learn nonsense.

Minimum dataset to get serious:

• Order created time, packed time, dispatched time, delivered time
• Location granularity (postal code isn’t enough for dense areas)
• Exception reasons (no access, customer unavailable, wrong address)
• Rider/vehicle type and shift windows

Step 2: Fix the high-cost exceptions first

Most last-mile cost blowouts come from a handful of recurring exceptions.

Start with:

• Redelivery loops
• Building access delays
• Cash-on-delivery friction (if applicable)
• Failed handoffs at peak hours

Then automate the decision: who should handle this job, with what constraints, and what fallback plan?

Step 3: Add automation where the environment is semi-predictable

Robots and autonomy work best where the “world” is consistent:

• Same routes daily
• Similar drop-off points
• Known pedestrian density patterns
• Clear geofenced operating zones

This is how you avoid expensive “hero projects.”

Step 4: Close the loop with monitoring and human escalation

Autonomy needs an operations layer:

• Remote monitoring for edge cases
• Clear escalation paths (“robot stuck”, “access denied”)
• Fast incident logging for retraining

A useful KPI set:

• Cost per successful delivery
• Successful deliveries per hour (robot vs human vs hybrid)
• Rescue rate (how often a human must intervene)
• Complaint rate by route and building type

People also ask: what does this acquisition signal for 2026?

Will delivery robots replace riders in Singapore?

No. The near-term pattern is hybrid fleets: humans handle variability and customer interaction; robots handle repeatable segments and low-contact routes.

Why acquire instead of build?

Because capability takes time. Infermove raised at least US$3.3 million before the acquisition, which suggests real engineering effort already sunk into data collection and training systems. Buying that compresses time-to-deployment.

What’s the “first-mile” angle?

First-mile is everything from merchant pickup to consolidation. Automation here reduces queueing and improves dispatch predictability—often a bigger win than the final doorstep moment.

What this means for your business (even if you don’t do delivery)

This story is a clean case study for AI mengoptimumkan laluan pengangkutan, automasi operasi, dan keberkesanan rantaian bekalan.

Three copyable principles:

1. Start with real-world data, not ideal workflows. Shadow systems beat whiteboard process maps.
2. Treat automation as a product with continuous training. If it doesn’t improve month-to-month, it’s just an expensive script.
3. Design for exceptions. The best AI operations teams spend disproportionate time on the 5% that breaks everything.

If you’re planning your 2026 operations roadmap, the question isn’t “Should we use robots?” It’s: Which parts of our logistics chain are predictable enough to automate—and which need better data before we even try?