Physical AI is the latest term for robot and drones with a mix of artificial intelligence (AI) that allows them to be smarter and more robust. The latest crop of physical AI devices integrates more sensors that feed more AI models to provide the robot with a better understanding of its environment. These sensors often include cameras and distance sensors like LiDAR and radar.
In this podcast, I talk with Yaniv Sulkes, Vice President of Physical AI at Hailo, about issues and solutions related to physical AI systems. We touch on some of Hailo’s AI hardware accelerators like the Hailo-8 and Hailo-10 (see figure). It supports large language models (LLMs) and visual language models (VLMs).
Hailo-10 delivers 40 TOPS of INT4 performance while consuming less than 2.5 W.
The Hailo-15 is a family of system-on-chip (SoC) AI platforms targeting imaging applications including physical AI. They incorporate an image signal processor (ISP) supporting advanced video analytics. It uses AI models to handle low-light denoising (<0.01LUX). All of this while consuming under 3 W of power.
Perceive, reason, and act are how people normally partition physical AI operation. The first is where Hailo SoCs come into play, allowing recognition of objects in its environment. This includes handling multiple camera data streams running AI models on this data. The Hailo-10 and Hailo-15 can also run more sophisticated models for reasoning.
