← all workDiDi · 2022–2024
Autonomous-Driving Model-Training Toolchain
Browser scene viewer + telemetry workspace for AD researchers — 3D point cloud, multi-camera, rosbridge live streams.
adviewer.PointCloud + Boxes
live · canvas / webglSimulated LiDAR rings + 3D bounding boxes (cars, peds) · ego vehicle wireframe centered.
What it is
The browser workspace that DiDi's AD researchers used every day — data, perception, planning, simulation. Replaced a fragmented mix of desktop ROS tools (RViz, Foxglove, PlotJuggler) with a single web-native viewer.
The interesting parts
- Three.js / WebGL scene viewer on top of RViz & Foxglove (Webviz) — full-fidelity LiDAR playback, multi-camera mosaic, 3D bounding boxes synced to perception outputs.
- ROS ↔ browser bridge via rosbridge — streamed rosbag2 recordings and live ROS topics into the browser.
- 200 Mb/s gRPC / Protobuf pipeline + multi-stream WebCodecs / H.264 video at <100 ms latency, 20 FPS.
- One rendering engine, three surfaces — the same code path drove React tools, a Vue analyst portal, and the Android in-vehicle HMI.
~200%render perf gain
<100 msvideo latency
>4 h/dayper-researcher use
Why it mattered
Became the AD data team's primary daily workspace. Earned the company-wide annual Efficiency Contribution Award.
The demo above is a stylized recreation — the internal viewer is not publicly accessible.