AI‑Era New Paradigm for Food‑Delivery Fulfillment: Insights from the D20 Smart Services Forum

Rider‑Centric AI Paradigm

Food‑delivery fulfillment now covers fresh food, apparel, medicine, flowers and 3C products, forming an online‑offline collaborative network among platforms, merchants, customers and riders. The design goal is to help riders earn more and improve their delivery experience.

Three‑Component Paradigm

The paradigm consists of GUI, multimodal Agent and smart‑hardware ZUI, which cooperate rather than replace each other.

GUI

GUI is the most mature component. It refines details for rider scenarios such as single‑hand operation on a phone mounted on a holder and click‑area design for waterproof cases.

Multimodal Agent

In the GUI the information architecture is predefined; in the Agent user input is free and the system behaves as a black box, producing dynamic responses. General large models perform poorly because they lack rider‑specific corpora (dialects, industry slang, noisy acoustic environments).

Problem 1 – Slow Conversation

Goal: natural, interruptible, single‑utterance multi‑task dialogue.

One‑Shot Wake‑up: compress “wake‑word + command” into a single utterance for single‑turn execution.

Barge‑in: allow riders to interrupt the system at any time.

Concurrent Intent Parsing: enable a command such as “confirm arrival, confirm pickup, then start navigation” to be parsed and queued.

Problem 2 – Misunderstanding Slang

Riders use compressed slang (e.g., “卡餐”). The team built a five‑step mapping “scenario → corpus → intent → match” to create a rider‑slang dictionary that translates slang into actionable intents (e.g., “merchant delay, report anomaly?”). For ambiguous expressions like “can’t find, how to get there,” the Agent now adds fuzzy‑intent collections combined with the rider’s GPS location and asks clarifying questions.

Problem 3 – Incorrect Decisions

Even when the Agent understands the request, generic model outputs may violate platform‑specific rules. The solution equips the Agent with two specialized skills: a deep knowledge of platform rules (Skill +) and multi‑layer dynamic strategy handling (Skill +).

Intelligent Proactive Agent

The Agent is extended from reactive Q&A to a proactive assistant that analyzes scenes, processes data and proposes actions for rider confirmation.

Example: the final 50 m of delivery (urban villages, office buildings, hospitals, schools) often causes riders to wait while other orders time out. By aggregating delivery‑point data, customer profiles, order notes, product attributes and spatiotemporal variables, the Agent can pre‑emptively contact customers, confirm delivery methods and push voice or HUD notifications to the rider, reducing wait time.

ZUI Smart Hardware

Smart hardware serves as the rider’s second pair of eyes. Phones require looking down, occupy a hand and perform poorly in rain. Alternative devices—helmet HUDs, drones, smart lockers, climbing robots, autonomous vehicles—can each host their own Agent.

Multiple hardware devices raise standardization challenges: defining HUD content, handling varied lighting conditions, ensuring readable font sizes at speed, and establishing visual hierarchy between general information and warnings. Systematic standards are needed to address these issues.

Conclusion

Integrating GUI, multimodal Agent and ZUI redefines fulfillment for riders, who constantly race against time.