How Baidu Achieved Unmanned Delivery with Risk‑Driven Testing and AI

Introduction

Baidu’s intelligent testing team has been exploring risk‑driven delivery as a core research direction. Three observations motivate this work: most projects generate no bugs, many test tasks fail to uncover defects, and testers often misjudge coverage.

Unmanned Delivery Concept

Unmanned delivery refers to a pipeline where testing and release steps require no human intervention. The article poses three common questions: what is unmanned delivery, how can it be achieved, and what benefits does it bring?

Evolution of Delivery Modes

Testing has progressed from manual to semi‑automated and finally to fully automated pipelines. In Baidu Search, DevOps and autonomous testing now handle over 90% of requirements, yet testers still spend considerable effort on failure analysis, manual verification, and communication.

Human Dependencies in the Current Pipeline

Decision dependence: Test tasks are selected by static configuration or manual decisions.

Process dependence: Multiple hand‑offs between developers and testers increase communication cost.

Conclusion dependence: Risk assessment relies on personal experience rather than automated analysis.

Eliminating these dependencies requires intelligent execution and risk‑assessment capabilities.

Three Core Capabilities for Unmanned Delivery

1. Complete Testing Capability

Testing must be both comprehensive and effective. For example, Baidu’s DIFF testing compares baseline and test versions but suffers from noisy results and manual interpretation. Baidu introduces “human‑like analysis” that combines white‑box data, models, and business logic to automatically judge whether a diff is caused by code changes and assess its business impact.

Two methods are used:

Knowledge‑base rules derived from accumulated business expertise.

Historical execution data to train a code‑field association model, which is applied online to infer causality.

2. Stable Build Capability

Frequent pipeline failures force testers to intervene repeatedly. Baidu adopts the stability standards used for online services to improve build reliability, focusing on three areas:

Dependency governance: Unified management of machines, code, data, and third‑party services.

End‑to‑end stability: Monitoring, early alerts, automatic error‑code handling, and self‑healing mechanisms.

Build digitization: Collecting and analyzing build metrics to quantify stability.

3. Precise Evaluation Capability

Baidu builds a quality‑score model that transforms delivery data (code metrics, developer profiles, test results, coverage, etc.) into features. Labeled historical data train the model, which outputs a risk score during admission/exit. The score is combined with business rules: low‑risk items are auto‑approved, high‑risk items trigger manual review.

Benefits of Unmanned Delivery

With comprehensive testing, stable builds, and automated risk evaluation, Baidu Search now achieves over 90% autonomous testing and more than 40% unmanned delivery. This frees testers to focus on higher‑value activities and reduces per‑requirement delivery costs.