Grab Deploys Delivery Robots to Accelerate Pickups

Grab is preparing to deploy an AI-enabled delivery robot, Carri, and showcased 13 AI features at GrabX 2026 as part of a broader push into robotics and embodied intelligence. CEO Anthony Tan framed Carri as an assistant for drivers, not a replacement, noting drivers currently lose around 10% of earning time locating restaurants or waiting for customers. The rollout follows Grab's January acquisition of Chinese robotics company Infermove, acquired to accelerate development of autonomous, manipulation-capable delivery robots.

Infermove's public profile and reporting describe a stack built for unstructured, last-mile environments using a mix of data-driven learning techniques. Key points practitioners should note:

• •Infermove emphasizes imitation learning, reinforcement learning, and proprietary end-to-end algorithms to handle mobile manipulation and navigation in cluttered pedestrian settings.
• •The company uses a crowdsourced data pipeline called Rider Shadow System that collects driving and interaction data from last-mile mobility devices, accelerating real-world training data acquisition and reducing reliance on purely simulated data.
• •Product-level capabilities described include sidewalk delivery robots with upper-limb manipulation and sensory stacks likely tuned for perception and contact interactions, enabling the robot to find a merchant location in a mall and transfer packages to a human rider.

This is not a theoretical research demo. Grab's strategy pairs software advances with operational scale across Southeast Asia, a market with dense urban malls, high pedestrian traffic, and diverse building typologies. Combining in-house logistics scale with Infermove's embodied intelligence addresses two persistent obstacles for physical robots: the sim-to-real gap and costly, slow data collection. The Rider Shadow System design is significant because it treats delivery riders as a distributed sensor network, turning routine rider trips into training data for complex manipulation and navigation behaviors. That approach short-circuits a common bottleneck for embodied AI startups and increases the probability of real-world robustness faster than simulated-only methods.

Deploying Carri as a driver aid preserves existing workforce economics while increasing throughput. If robots can reliably shave wait and search time, platform-level metrics such as deliveries per driver-hour, average order cycle time, and unit economics per order will improve. Reuters and industry commentary have flagged potential cost reductions in last-mile delivery from robots and drones, with some estimates suggesting per-order costs could fall dramatically in high-density corridors.

Implementation details matter. Monitor pilot geographies, sensor suites, human-robot handoff success rates, regulatory responses in Singapore and regional markets, and whether Grab open-sources tooling or integrates Infermove tech into fleet teleops. Also watch metrics Grab reports for driver earnings, delivery latency, and robot uptime during the early pilots.

Grab is moving from acquisition to deployment, combining Infermove's embodied learning techniques with operational scale. For practitioners, this is a live example of how crowdsourced, real-world training data and hybrid learning methods are being used to solve the last-mile robotics problem at scale.