Baidu’s Open Source Self-Driving Platform Apollo in Detail

Mario Herger, , ,

The first presentation outside China of Apollo, the open source platform for the development of self-driving cars by the Chinese internet behemoth Baidu, took place yesterday in Baidu’s Silicon Valley-offices in Sunnyvale.

In front of several hundred participants Baidu talked about the motivation, functionality, architecture, partnerships and showcased several projects with partners. The official launch of Apollo 1.0 was July 5th 2017 at the Baidu Create AI conference. The source code, which can be found on GitHub, was downloaded 6,000 times and has seen 1,300 forks.

Baidu aims at democratizing autonomous driving, not monopolizing it. According to Baidu president Ya-Qig Zhang the company believes in the open source approach to hand over control to developers and accelerate the development of this technology.

It’s obvious that Baidu aims at closing the gap to companies such as Google and boost its own cloud services. Interesting to see is that Google used a similar tactics ten years ago against Apple, when it open source its mobile OS Android. Now Baidu finds itself in this role against Google.

The Apollo-framework is composed of four layers, starting with the reference vehicle platform, on which the reference hardware  layer resides. On top comes the open source platform and the cloud service layer.

Apollo_Layers

The roadmap for version 1.5 and higher is pretty aggressive.

  • July 2017: closed venue driving
    • Control
    • Localization
    • Runtime Framework
    • Reference Hardware
    • Reference Vehicle
    • Data Platform 1.0
    • Labelled 3D obstacle data
    • Road Hackers data
    • HD map technology cooperation
  • September 2017: fixed lane driving
    • Obstacle Perception (Velodyne 64 Lidar, CUDA, CuDNN, Caffe, NVIDIA GPU 10Hz)
    • Planning (Multi-dimensional inputs, traffic law plug-in, DP/QP path and speed optimizer)
    • HD Map
    • Simulation Service
    • End-to-End Learning (Lateral: curvature; longitudinal: velocity and acceleration; FCNN, Convolutional LSTM)
    • Reference Hardware-Lidar
    • Artificial road scene data
  • December 2017: simple urban road driving
    • Advanced Perception
    • Security Service
    • Data Platform 2.0
    • Labelled obstacle behavior data
    • Labelled 2D obstacle data
    • Actual road scenario data derived from driving logs
  • December 2018: geo-fenced highways and urban roads
  • December 2019: highway and urban road alpha version
  • December 2020: citywide highway and urban roads

In 2019 Level 3 autonomous driving is target, and for 2021 Level 4. So far Baidu has signed 70 partnerships globally with OEMs, Tier 1 and 2 suppliers, as well as startups. An increase of 40 percent from July. Each week updates are released on GitHub, and more hardware and car model support is added.

Videos with partner showcases included the Chinese bus manufacturer King Long and the university spinoff Momenta.

Momenta

A larger focus is on maps. Those help the car to localize itself, with perception of objects around itself, and planning the route. Maps also help developers in software development. The maps include traffic lights, their precise positions and heights, and related lane markings. Also virtual lane markings are included, like the ones required at intersections to help direct the car. Additional map elements are lane dimensions, speed limits, or speed bumps.

Until 2020 Baidu aims to have all Chinese roads and highways in the maps. Other countries can be integrated through open standards such as OpenDrive.

An important aspect was the development of a simulator. To get to mass production, autonomous vehicles must drive 10 billion kilometers for proof of concept. With 100 vehicles driving 24 hours seven days a week , this would still mean 225 years of driving. Simulators can accelerate that and reduce costs. Both Uber and Waymo recently gave a glimpse into the workings of their simulators, with which they “drive” billions kilometer per year.

Baidu also announced an investment fund with 1.5 billion dollars. The money will be going to 100_ project in the next three years, focusing on startups working on hardware, product, data, and developer ecosystems.

Besides funding open source developers also requested more training data, support with testing and working with regulators, as well as additional hardware and vehicle model support. Baidu is working on supporting that as well.

This article has also been published in German.

Published by Mario Herger