Fully autonomous vehicles require a strong sensing system in order to capture 360-degree environmental information. Moreover, multiple sensor types are needed for redundancy and to enhance the information quality.

These multiple, heterogeneous data streams must then be merged, in order to obtain a single, more precise representation of the environment. The challenge is that this sensor fusion process is generally a resource-intensive task.

The computational burden presented by sensor fusion is worsened by the real-time planning capabilities that are mandatory for an autonomous vehicle. Indeed, the required computational resources are typically achieved through the use of high-power computing systems, such as GPU- or CPU-based architectures. However, these brute-force processing architectures generate substantial amounts of heat that can only be dissipated through liquid cooling, which opens another wide set of challenges, including higher costs and larger form factors. Additionally, traditional computing systems have high power consumption, which is particularly important for electric vehicles where this consumption results in larger batteries and weight that reduce range while further increasing costs This presentation will examine an alternative implementation of an autonomous driving system, based on a low-power, single-chip architecture. Videos will be shown from recent road tests of this vehicle, featuring single-chip L4 driving, which uses data acquired from 18 cameras and nine radars.