USE Network launch I UAV Works VALAQ l Cable harnesses l USVs insight l Xponential 2020 update l MARIN AUV l Suter Industries TOA 288 l Vitirover l AI systems l Vtrus ABI

8 Platform one A key autonomous controller for driverless cars is getting an upgrade to include a machine learning processor (writes Nick Flaherty). NXP is working with French start-up Kalray to develop safe, reliable and scalable solutions for autonomous driving, combining NXP Automotive solutions and Kalray’s MPPA (Massively Parallel Processor Array) intelligent processors. The MPPAs will be used in the next generation of the BlueBox autonomous driving reference platform alongside NXP’s S32 family of safe automotive processors and automotive-grade Layerscape network and comms processors. The resulting system will support designs from Level 2 (partial driving automation) to Level 5 (full vehicle automation) with common hardware and software. This will combine CPU processing, neural network computing, functional TDK has launched a high accuracy six- axis motion sensor in a single package to simplify IMU design (writes Nick Flaherty). The InvenSense ICM-42688-P sensor combines a three-axis gyroscope and a three-axis accelerometer in a 2.5 x 3 x 0.9 mm package for an IMU in autonomous ground and air systems. Compared to traditional consumer sensors, the sensor has a 40% lower noise figure, of 2.8 mdps/Hz for the gyroscope and 70 µg/Hz for the accelerometer, and twice the temperature stability, making it suitable for autonomous designs. This improvement comes from using a new clock input that reduces timing errors by using a 31-50 kHz clock and a safety capabilities and an optimised tool chain for automated driving. Kalray’s third-generation MPPA3 processor, called Coolidge, is optimised for computer vision and machine learning in autonomous vehicles. It combines 80, 64-bit processor cores, each with its own AI accelerator core. This is quite different from the second- generation processor, the MPPA Bostan, which is a 288-core graphics processor. high-resolution 16-bit analogue-to-digital converter. This combination, along with programmable digital filters, provides an 8x increase in gyroscope resolution to 19 bits and a 4x rise in accelerometer resolution, to 18 bits. The data is stored in 20-bit format in a 2 kbyte ‘first in, first out’ memory to maintain the accuracy. The sensor has I3C, I2C and SPI serial comms links. There are two programmable interrupts to minimise system power consumption, with a low- power mode that consumes 0.88 mA. It also supports a VDD operating range of 1.71-3.6 V and a separate digital I/O supply of 1.71-3.6 V. A development kit, the DK-42688-P, includes MotionLink software that captures and shows the sensor data. Memory bandwidth is also vital for machine learning, so the 80 cores in the Coolidge chip are connected by a 600 Gbyte/s network-on-chip with high- speed PCIe Gen4 to external DDR4 memory and 200 Gbit/s Ethernet to the rest of the BlueBox system. This provides a performance of up to 1.1 TFLOPS and 25 TOPS with support for complex data flows. The network-on- chip allows the array of processors to be segmented so that separate applications can run securely in different parts of the chip, providing spatial isolation. That means safety-critical applications can be locked into specific parts of the chips, while other, less critical applications can run in other parts. There are dedicated low-power DDR channels to each of the clusters. This also means a real-time operating system can run alongside a general- purpose operating system such as Linux. It also includes an embedded motion driver, which is a set of APIs to configure various aspects of the platform including ICM-42688 sensor parameters such as full-scale range, output data rate, low- power or low-noise mode, and the I3C/ I2C/SPI interface to the host controller. Driverless cars Navigation Machine learning upgrade Six-axis motion sensor June/July 2020 | Unmanned Systems Technology The BlueBox will feature intelligent processors Putting the gyroscope and accelerometer in a single package simplifies IMU design