Baidu unveiled an open source “OpenEdge” edge computing platform and an open Linux-based “Apollo 3.5” autonomous car stack. OpenEdge dev boards include an Intel-based BIE-AI-Box in-car visual analytics board and NXP-based BIE-AI-Board for IoT.

Baidu, which is often referred to as the Google of China, has announced an open source, AI-infused OpenEdge edge computing platform with development boards based on Intel and NXP SoCs. The news follows Baidu’s announcement earlier this week that it was releasing version 3.5 of its open source Linux-based Apollo self-driving software stack, as well as a new Apollo Enterprise platform based on it designed for vehicle fleet management (see farther below). The open source platforms were announced at this week’s CES show in Las Vegas.

Apollo 3.5 test car detail view
(click image to enlarge)

 
OpenEdge

The sketchy details on OpenEdge included no mention of Linux, but judging from what we know about the first development boards — an Intel-based BIE-AI-Box in-car camera analytics board and NXP-based BIE-AI-Board — Linux is a very likely suspect.

According to several sources, including VentureBeat, OpenEdge is the local package component of Baidu’s commercial Baidu Intelligent Edge (BIE) platform. Much like cloud-driven edge analytics platforms such as Google’s Cloud IoT Edge platform, OpenEdge is designed to collect data, distribute messages, perform AI inference, and synchronize with the cloud. The idea is to reduce latency and networking costs while making edge IoT systems and gateways more resilient to the loss of an Internet connection.

Like BIE, OpenEdge is a cloud-based management suite for managing “edge nodes, edge apps, and resources such as certification, password, and program code,” says VentureBeat. It supports models trained on AI frameworks like Google’s TensorFlow and Baidu’s PaddlePaddle. Edge devices running OpenEdge will be able to cache data and perform on-device processing if the cloud connection is lost.

According to TechRepublic, OpenEdge will seamlessly exchange data with Baidu ABC Intelligent Cloud and perform filtering calculation on sensitive data. It will also cache data, perform standalone computing, and provide real-time feedback control.

Intel’s BIE-AI-Box is an in-vehicle video analytics computer that runs OpenEdge on an unnamed processor. The development platform “connects with cameras for road recognition, car body monitoring, driver behavior recognition, and other tasks,” says VentureBeat.

According to TechRepublic, BIE-AI-Box connects with cameras in vehicles to optimize video analysis. The board will provide “AI apps for road recognition, car body monitoring, and driver’s behavior recognition,” says the story.

The BIE-AI-Board, meanwhile, is a lower-power solution based on an NXP SoC — probably one of its i.MX8 family of Arm-based SoCs — designed for cameras, drones, robots, and other IoT hardware. Early partners have integrated the BIE-AI-Board within electric vehicles to “assess the health of chargers and with agricultural drones to analyze crop spectral data,” says VentureBeat. It was unclear if the hardware would be as open as the OpenEdge software.

 
Apollo 3.5

The latest Apollo 3.5 autonomous vehicle software platform goes beyond other driverless car stacks in its ability to “perform in complex urban and suburban driving scenarios,” says Baidu. The open source, Linux-based software competes with Nvidia’s newly announced, Xavier based Drive AutoPilot as well as Tier IV’s Autoware. In addition to these Linux-driven platforms, Apollo competes with proprietary hardware/software driverless car projects including Waymo.

Apollo 3.5 architecture (left) and Ubuntu-based development screen
(click images to enlarge)

Apollo 3.5 includes features such as algorithms for handling unprotected turns, speed bumps, clear zones, side passes, narrow lanes, and parking, says VentureBeat. The Apollo 3.5 GitHub page states: “Apollo 3.5 is capable of navigating through complex driving scenarios such as residential and downtown areas. The car now has 360-degree visibility, along with upgraded perception algorithms to handle the changing conditions of urban roads, making the car more secure and aware. Scenario-based planning can navigate through complex scenarios including unprotected turns and narrow streets often found in residential areas and roads with stop signs.”

This latest Apollo version integrates an “entirely open source” Cyber RT runtime framework. Apollo 3.5 also adds support for the latest lidar sensors, such as Velodyne’s VLS-128, which is used in its Alpha Puck lidar device.

Apollo 3.5, which now comprises 400,000 lines of code, runs on an Apollo Kernel based on Linux 4.4.32 within Ubuntu 14.04. It’s being “tested, contributed to, or deployed” by Intel, Nvidia, NXP, and over 130 global partners, says VentureBeat. This is interesting considering that Nvidia is a competitor, but then again, everyone is leaving their options open at this point.

Apollo 3.5 hardware block diagram and test car
(click images to enlarge)

The new release follows the Apollo 2.0 platform announced a year ago at CES, which is designed for simpler urban-driving environments. In April, Apollo 2.5 added “improved vision-based perception, real-time relative mapping, new driving scenarios, and visual debugging tools,” says VentureBeat.

This was followed by a 3.0 release with new support for valet parking, autonomous mini buses and microcars, and integration with Baidu’s voice-activated telematics software, “which can perform facial recognition and monitor drivers for signs of fatigue,” says the story. It also added Responsibility-Sensitive Safety (RSS) technology from Intel subsidiary Mobileye. A fully autonomous Apollo design for all driving conditions is due in 2020.

Apollo 3.5 IPC (Neousys Nuvo-6108GC) and Baidu sensor fusion unit
(click images to enlarge)

The Apollo stack runs on an Apollo IPC system, which is a modified version of the Neousys Nuvo-6108GC industrial PC. It provides an Intel C236 Platform Controller Hub with support for either an Intel Xeon E3 v5 or 6th-Gen Core i7/i5 CPU. On the Apollo IPC platform the system is equipped with Asus GTX1080 GPU-A8G-Gaming GPU card via a PCIe x16 PEG interface.

The 360 x 174 x 164mm Nuvo-6108GC system supports up to 32GB DDR4 and offers 2x GbE, 4x USB 3.0, 2x DVI-D, and 2x serial ports. In addition to the PEG port, the system provides 2x PCIe x8, mini-PCIe, M.2 B-key, and SIM slots. Four SATA bays are available, one of which is loaded with a 1TB HDD on the Apollo IPC. The -25 to 60°C computer offers 1 Grms, 5-500 Hz vibration resistance.

The Apollo hardware development platform also includes a Baidu-built sensor fusion unit. In addition, there are camera, lidar, radar, and navigation systems, among other components.

 
Apollo Enterprise

Baidu also announced an Apollo Enterprise fleet management platform based on Apollo 3.5 Apollo Enterprise will include a la carte solutions for highway autonomous driving, autonomous valet parking, fully autonomous mini-buses, and an intelligent map data service platform, according to a TechCrunch story. It also includes Baidu’s DuerOS voice assistant.

Apollo Enterprise is already being used by Hyundai Kia, Ford, and VM Motori, in addition to Chinese car manufacturers Chery, BYD Auto, and Great Wall, says VentureBeat. In addition, over 100 auto OEMs are said to have adopted Enterprise on 300 car models.

Finally, Baidu said it is open-sourcing its vehicle-to-everything (V2X) Apollo Intelligent Vehicle Infrastructure Cooperative System platform, including software and reference hardware. This will be used in an autonomous taxi trial in Changsha, China, which will involve 100 robotaxis operating on 130 miles of city roads equipped with Baidu’s V2X infrastructure.

 
Further information

Baidu’s Apollo 3.5 is available now, and the hardware development platform components are available to partners with undisclosed pricing. There was no mention of a ship date or website for OpenEdge and the BIE-AI-Box and BIE-AI-Board dev boards. More information on Apollo may be found on Baidu’s Apollo website and GitHub page.