The Apollo community has built a machine-learning based auto-calibration system for autonomous driving vehicles. By August 2018, the system had been tested on more than two thousand hours with around ten thousands kilometers’ (6,213 miles) road tests and has proven to be effective. The system is automated and intelligent, due to which, it is suitable for mass-scale self-driving vehicle deployment.
Why was Apollo auto-calibration system introduced?
Following are the main issues that the current system faces:
- Manual calibration is time consuming and error prone: The performance and safety of an autonomous driving vehicle depend on its control module. This module includes control algorithms that require vehicle dynamics as input and then sends command to manipulate the vehicle. Performing this calibration in real-time is difficult, that is why, most of the research-oriented autonomous vehicles do manual calibration in one-by-one fashion. Manual calibration consumes a lot of time and is prone to man-made mistakes.
- Variation in vehicle dynamics: While driving the vehicle dynamics change (i.e. loads change, vehicle parts will be worn out over time, surface friction), and manual calibration cannot possibly cover them.
How does Apollo auto-calibration system work?
The auto-calibration system depends on the Apollo control module, which consists of an offline model and online learning algorithm
First, a calibration table is generated based on human driving data that best reflects vehicle longitudinal performance at the time of driving. It performs three functions:
- Collects human driving data
- Preprocesses the data and select input features
- Generates calibration table through machine learning models
The online algorithm updates the offline table based on real-time feedback in self-driving mode. It tries to best match the current vehicle dynamics based on offline model established from manual driving data. It performs the following functions:
- Collects vehicle status and feedback in real time
- Preprocesses and filter data
- Adjusts calibration table accordingly
To know more details on how this model works and helps to solve the manual calibration problem, check out their published paper: Baidu Apollo Auto-Calibration System – An Industry-Level Data-Driven and Learning based Vehicle Longitude Dynamic Calibrating Algorithm.
Apollo 11 source code: A small step for a woman, and a huge leap for ‘software engineering’
Baidu open sources ApolloScape and collaborates with Berkeley DeepDrive to further machine learning in automotives
Tesla is building its own AI hardware for self-driving cars