Issue 54 Uncrewed Sytems Technology Feb/Mar 2024 uWare uOne UUV l Radio and telemetry l Rheinmetall Canada medevacs l UUVs insight DelltaHawk engine l IMU focus l Skygauge in operation l CES 2024 report l Blueflite l Hypersonic flight

10 February/March 2024 | Uncrewed Systems Technology Provizio in Ireland has developed an imaging radar able to provide data from an area of 1 km2 around a vehicle (writes Nick Flaherty). The Prime radar provides a dense point cloud of data, created with generative AI, to replace a laser Lidar sensor in autonomous vehicles, starting with mining and agricultural applications. This uses a commercial radar chip from Texas Instruments and a patented, active antenna using chips designed in-house in the transmit-and-receive path in front of the radar chip. “We have two vehicles on the road in Ireland with a 1 km2 cocoon,” said Barry Lunn, CEO and founder of Provizio. “One is testing the forward-facing and the other the L3+ rear-facing mode with over 600 m range.” “That’s what allows us to get the extended range and resolution as we are adjusting the phase and time to illuminate a wider aperture,” Lunn explained. At the radar’s front end is a planar antenna, with the Provizo-designed chips in front of the radar chip. These operate at 76-81 GHz and are built at Global Foundries. On the receive path, the chip is a low noise amplifier (LNA). This reduces the noise floor, rather than increasing power as that would add noise. The radar front end is just one part of the design, which includes software to that cause problems, such as side lobes, they are experimenting with reducing at the signal level and the output.” “We use generative AI to improve the point cloud by 4-10%, but that’s not improving the resolution – that’s improving the inferencing performance post-DSP,” Lunn pointed out. “What we are doing is training lower-grade radar on higher-grade data, and training radar on a Lidar point cloud with classification.” Mining and agriculture applications will be the first to adopt the technology. “By 2028, they want the entire fleet automated with L4 and L5. We have data from mines for SLAM odometry, but we showed we could classify pedestrians in the dust and rain while providing the odometry,” Lunn added. The company plans to start production of the radar with Mergon in the next few months. Radar Imaging radar’s safety cocoon produce a point cloud similar to a Lidar laser sensor. This is fed into a data-fusion algorithm with data from a camera, and it works with processors from Nvidia and other suppliers. “Our point cloud is 40-50x denser than a standard antenna for the perception and classification,” said Lunn. “We made everything modular, so an established manufacturer can use the active antenna and the chips with their own radar to get better resolutions.” The key is the expertise in machine learning (ML) combined with more traditional digital signal-processing (DSP) algorithms. “The really important thing is the fact that we have a perception team gives us a huge advantage in the design of the radar for perception rather than for detection. As the ML team gets deeper into the stack they want raw data. Some of the things The Prime radar provides a dense point cloud of data, created with generative AI Console showing automatic lane change initiated