Apollo 5.0 — Geo-Fenced Autonomous Driving Volume Production
This is the part one of the two-part technical blogs that walks you through all the newly released features and tools in Apollo 5.0. Part 2 focuses on the existing module upgrades, read it here.
During CES 2019, Baidu Apollo released a milestone version — Apollo 3.5, which enabled developers to achieve urban autonomous driving. This is the only open-source solution in the industry that empowers developers around the world to handle complex driving scenarios such as unprotected turns, speed bumps, clear zones, side passes, navigating narrow lanes, and autonomous parking, as well as seamless autonomous driving in downtown and residential areas.
Fast Forward to July 3rd, the Apollo Team is excited to introduce Apollo 5.0 — an upgraded version that is packed with even more features and offerings to accelerate volume production for geo-fenced autonomous driving in a complex urban environment. Without further ado, let’s dive in!
Apollo Data Pipeline
While many of Baidu Apollo’s developers and partners are aiming for volume production and fleet scalability, the biggest obstacle in their paths right now is data — extraction, processing, utilization, and conversion of raw sensor data into more meaningful data that can ultimately enhance the vehicle’s performance. In a 2018 report made by Accenture, we see some key findings regarding data:
“Test AVs today generate between 4 and 6 TBs of data per day, with some producing as much as 8–10 TBs depending on the number of mounted devices and their resolution. By comparison, the typical person’s video, chat and other internet use averages about 650 MB per day. That means, on the low end, the data generated from one test car in one day is roughly equivalent to that of nearly 6,200 internet users.”
This is an enormous amount of data! Combining with the use of efficient cloud services and accurate AI models, different algorithms’ capabilities and accuracies can be trained. But the core challenge is separating meaningful data from the raw sensor output. In Apollo 5.0, we brought you the Apollo Data Pipeline. It consists of a combination of open-source tools and service offerings that helps developers tackle the data challenges highlighted above in autonomous driving commercial deployment.
Smart Data Recorder
In the old days, the only way to get meaningful data was to analyze the whole sha-bang of sensor raw data. This method was extremely demanding on the data storage mechanisms, extremely time-consuming and complex extraction methods, and very painful to transfer. This process is not ideal, especially when managing large fleets where the majority of the driving data is unnecessary to record. You might encounter a handful of situations that need comprehensive data recording that would need further pre-processing. Apollo 5.0 has the solution to this unique requirement by way of the Smart Data Recorder. The Smart Data Recorder empowers you to capture drive events that are specifically tailored to your requirements with a click of a button in our HMI tool Dreamview. With Smart Data Recorder, you can reduce the amount of data storage by 90% by only capture the meaningful data.
Large-scale Cloud Processing
Data can be used to improve all aspects of autonomous driving software, from map & localization, perception, planning to control and many more modules. But in order for data to help enhance your algorithms and software performance, it needs to be processed correctly. That’s why the Apollo Team has built a Large-scale Cloud Processing Engine which helps train large datasets to generate data models that can be used to enhance your autonomous driving software capability in a scalable way. We are excited to introduce the first application that was built on top of this powerful cloud processing engine — Vehicle Control Calibration Service.
Vehicle Control Calibration Service
Prior to Apollo 5.0, the only way to perform vehicle control calibration was by using an open-source tool that generates a control parameter table. The new Vehicle Control Calibration Service includes an in-vehicle data collector, a task progress visualization tool and real-time data verification. In the process of massively parallelizing cloud training, the tool performs a sanity check on the uploaded data, and then proceeds with data cleaning, filtering and preprocessing. After that, the AI-based module starts to perform full data modeling which greatly improves the accuracy of the final result. Before notifying the users, the results would go through an automatic testing and verification process and the model’s performance would be analyzed. Users then receive a comprehensive report via an email. Compared to the previous versions of the tool, the amount of vehicle control calibration data has multiplied 4 times, the efficiency of control calibration has multiplied 16 times and the accuracy of control calibration has increased by 26%. This service is able to perform control calibration on multiple vehicles simultaneously by achieving parallelism, which drastically increased the scalability of the control calibration service. This service is currently in the beta testing phase. If you are interested in becoming our beta user, please send us a request at apollopartner@baidu.com.
Control-in-the-Loop Simulation
Traditional dynamic models are often modeled using a rule-based approach — simple and complex. Such traditional modeling methods have various limitations in terms of model complexity, lack of clarity, versatility, and scalability in terms of development costs. The machine learning-based dynamics model based on the Apollo Data Pipeline has many advantages: the model is extremely sophisticated, it is safer to deploy in real vehicles especially you are able to test your parameters virtually. This process speeds up the development time as you have already caught the module’s errors in the test and are able to make it more efficient prior to road tests. Beyond those advantages, the versatility and scalability of this new model are unparalleled compared to other rule-based models. With the Control-in-the-Loop Simulation feature enabled, it is now possible to gauge your actual control parameters on your simulation platform without having to test with a real vehicle. This service is currently in the beta testing phase. If you are interested in becoming our beta user, please send us a request at apollopartner@baidu.com.
Create Customized Scenarios with Apollo Simulation Platform Dreamland
In Apollo 5.0, the Apollo Team implemented one of the most requested features for the simulation platform Dreamland — introducing Scenario Editor. This highly desired feature allows developers to create and customize their own scenarios within the easy-to-access simulation platform to validating their algorithms and increase the efficiency of the development cost. To avoid any dependencies in the development environment, the Scenario Editor was built on the cloud and can be accessed via any browser of your choice. In addition, the new editor offers a wide variety of scenario elements (such as maps, ego-vehicle, traffic lights, pedestrians, and etc.) that empowers developers to create diverse scenarios that provide an accurate and customized testing environment. Beyond all these features, Scenario Editor also comes with a set of grading metrics that you can select to evaluate your software performance against expected rider experience as well as the traffic regulations. This versatile tool is live on http://azure.apollo.auto, and it’s free to access to all Apollo developers. We even made a quick tutorial video to show you how easily you can create your own scenario! Check it out.
Apollo Synthetic Dataset
To empower developers to fine-tune their Perception algorithms, the Apollo Team released the largest synthetic dataset with diverse high-fidelity scenes and comprehensive ground truth to help developers with their autonomous driving R&D. This dataset contains more than 270,000 photo-realistic distinct images from various virtual scenes of high visual fidelity, including highway, urban, residential, downtown, indoor parking garage environments. These virtual worlds were created using the Unity 3D game engine. The biggest benefit that this synthetic dataset provides is the precise ground truth. Another benefit is it also contains diverse environmental variations, such as different times of day, different weather conditions, different traffic/obstacles, and varied road surface qualities, which are relatively harder & costlier to achieve in the real world. Furthermore, Apollo Synthetic Dataset provides an extensive set of ground truth data: 2D/3D object data, semantic/instance-level segmentation, depth, and 3D lane line data. This dataset is freely open to everyone, and download it today at http://apollo.auto/synthetic.html.
As you can see, Apollo 5.0 packs with loads of exciting new features and developer-friendly tools. Everything is live and ready for access today on our GitHub page and website. To make everything easier for you, we summarized a list of links for all the new features and tools you read above. We hope you like this milestone release, and join us in the smart mobility revolution!
Apollo 5.0 GitHub Release Page: https://github.com/ApolloAuto/apollo/releases/tag/v5.0.0
Vehicle Control Calibration Service (Beta): Email us at apollopartner@baidu.com to join our Beta Program
Control-in-the-Loop Simulation (Beta) & Scenario Editor on Dreamland: Create an account at http://azure.apollo.auto and send an email to idg-apollo@baidu.com.
Apollo Synthetic Dataset: http://apollo.auto/synthetic.html
