Luxeed G9 L3 Testing: The Tug-of-War Between Open Source 'Transparency' and Closed Systems

Huawei Smart Selection Car Product Director Peng Lei test-drove the Stelato G9 L3 functional test vehicle and described the experience as "epoch-making" — this statement sparked a subtle question in the open-source community: when "epoch-making" technological breakthroughs happen within a closed supply chain, how do we verify if they are truly trustworthy?

As a long-time contributor to the autonomous driving open-source ecosystem, what I see is Huawei's contradiction regarding testing transparency: on one hand, there are public road tests, claimed videos, and topic generation; on the other hand, core algorithms, training data, and scenario libraries remain black boxes. This "epoch-making" experience looks more like a limited-edition demo than a reproducible engineering achievement from the community's perspective.

From "Epoch-Making" to "Reproducible" — The Transparency Issue of L3 Testing

The "epoch-making" experience mentioned by Peng Lei likely stems from Huawei's breakthroughs in end-to-end perception and integrated planning/control. However, the standard for measuring technical value in the open-source community has never been "how good it is," but rather "how easy it is to verify." A typical open-source project, such as the Carla simulator, includes detailed benchmark results and reproduction steps with every release.

Dimension	Open-Source Autonomous Driving Projects (e.g., Apollo)	Huawei Stelato G9 L3 Testing
Public Test Data	Some datasets open-sourced, standard test sets available	Not disclosed
Algorithm Reproducibility	Complete code and dependencies provided	Closed
Safety Assessment	Community crowdsourced testing, Issue tracking	Internal verification
Failure Case Sharing	Unified Crash record library	Not disclosed

Huawei's L3 test vehicles hitting the roads with labels is commendable as a "roadshow," but it lacks a community-participatory verification mechanism. On GitHub, we are used to using the compare view to contrast performance differences between commits, whereas Huawei's "epoch-making" experience currently relies solely on the product director's test drive notes.

[!info] Community Perspective

A truly "epoch-making" technology should provide a reproducible testing environment. Huawei could consider opening up parts of its scenario library, just as Apollo opened up the "Scenarios" folder, allowing developers to run tests under the same conditions instead of only showing edited videos.

Huawei's Test Vehicles: Black Box or Open-Source Reference Implementation?

From the exposed videos, the Stelato G9 test vehicle is equipped with LiDAR, millimeter-wave radar, cameras, and other sensors, printed with the words "L3 Level Autonomous Driving Road Test." This reminds me of Waymo's early Open Dataset on GitHub and Tesla's hardware reference designs. Huawei's hardware solution is likely self-developed, but is it possible that the software layer borrows from open-source community practices?

Imagine the structure of an open-source reference implementation:

StelatoG9_L3_architecture/
├── perception/
│   ├── lidar_processing (Based on PointPillars)
│   ├── camera_fusion (Based on BEVFormer)
│   └── radar_filter (Based on Kalman)
├── prediction/
│   ├── trajectory_forecast (Based on Transformer)
│   └── interaction_model (Based on Social LSTM)
├── planning/
│   ├── behavior_planner (Based on Rules + Learning)
│   └── motion_controller (Based on MPC)
└── system/
    ├── safety_monitor (Based on Formal Verification)
    └── hmi_interface (Based on ROS2 messages)

If Huawei could open-source this architecture in a modular way, even providing interface specifications for some components, it would greatly advance the community discussion on L3 safety. But the reality is that Huawei chose a closed route: all code runs on their proprietary MDC computing platform, exposing no APIs to external developers. This turns the "epoch-making" experience into a "miracle inside a black box," which the community cannot improve via Pull Requests.

Community Governance Perspective: Who Reviews the "Pull Request" for L3 Cars?

L3 autonomous driving means the system assumes full responsibility under specific conditions; if an accident occurs, liability attribution is key. In the open-source community, we ensure code quality through Review mechanisms and CI/CD pipelines. But automotive safety differs from software bugs — a wrong merge can lead to personal injury.

Huawei's governance model is "centralized": all decisions are made by internal expert teams, and consumers passively accept them. The open-source community advocates for "distributed" governance: every contributor can propose improvements, ensuring quality through Peer Review and test coverage. For L3 cars, this community governance model is clearly unrealistic, but its "transparency" principle can be borrowed.

For example, can Huawei publish the L3 system's "Safety Redundancy Design Document" and "Failure Mode Analysis"? Can it regularly issue "Security Advisories" describing known vulnerabilities and fixes, similar to the Linux kernel? These practices don't require open-sourcing the code, but

Original Link: https://www.ithome.com/0/980/073.htm