Multiple large-scale trials of autonomous robotaxis are currently taking place across several cities in the US, notably in Austin, Phoenix, Las Vegas, San Francisco and Atlanta. These are run by companies include Tesla, Zoox and Waymo (among others).
With such competition, is it better to build for accuracy? Or for cost and speed of roll out. Currently both are being trialled and it’s likely that only one architecture will survive.
In this blog we’ll give an overview of the two architectures, their relative strengths and weaknesses, and examine how LiDAR can be implemented more cost effectively and power efficiently to enable enhanced safety in such a competitive environment.
The sensor platforms
Let’s start with the cars themselves.
Waymo’s is arguably the most developed system and its latest (sixth-generation) vehicles implement a platform based on 13 cameras, four LiDAR sensors and six radar sensors in order to see up to 500 meters away. This is complemented by an array of external audio receivers to detect upcoming emergency vehicles and other audible signals, and data from all sensors are fed to an onboard AI to undertake navigation, steering and braking. Crucially, these are retrofitted to cars made by other companies.
Amazon’s Zoox increases the number of sensors, with eight LiDAR, 18 cameras, four additional thermal cameras and 10 radar units plus eight microphones. Similar platforms are also used the major Chinese analogs: Baidu, Pony.ai, AutoX, and WeRide.
Figure 1: LiDAR unit embedded above the wheel arch of an autonomous robotaxi
Taking the alternative approach is Tesla.
Like Waymo, Telsa builds in an array of vision cameras, in this case 8, to give a 360o view. But unlike Waymo, it uses its own cars with the sensors already integrated and it does not use radar or LiDAR. LiDAR in particular comes with significant costs and power consumption – for example LiDAR costing roughly $12,000 per vehicle, vs cameras, which cost around $400 per car. The lack of it therefore gives Tesla a significant cost advantage over its rivals, with its robotaxis being built for as little as $22,000.
The cost implications seem to be noted and Waymo’s latest vehicle has dropped the number of LiDAR units (its fifth-generation model used five), with costs cited as the reason for the reduction.
Tesla also maintains that using just camera is the most “human” way to approach self-driving, comparing it to people using just their eyes to navigate the road. The company also insists that LiDAR is “expensive and unnecessary.”
Is LiDAR Essential?
For safety, we’d argue yes, it is.
And we’re far from alone in that conclusion, with the US automaker Rivian openly criticising camera-only strategies in its AI and Autonomy Day, held in December. As Rivian put it to Drive Magazine: “in night conditions and fog [cameras] don’t perform as well as active light sensors like lidar… Lidar can actually double visibility at night. Cameras are great at seeing the world until they can’t.”
It’s notable that (in this initial phase, at least) Robotaxis tend to only operate only in cities in the Sunbelt. These have warm dry climates that are as simple to drive in as possible Phoenix, for example, has over 330 dry days per year with just over 7 inches (18 cm) of rain per year.
Figure 2: Map of US sunbelt (reproduced with permission under creative commons license, original here.)
Cameras work really well in such environments, they work less well in heavy rain, snow and fog and the rest of the US (and the world) is not so fortunate with the weather. As robotaxis start to be used in Europe, and other parts of the world that are less ideal, implementing a wide range of sensors becomes essential.
And even in these perfect conditions, cameras can fail with Bloomberg reporting a crash caused by the setting sun blinding the cameras of an autonomous vehicle in self-driving mode. This would not have happened with LiDAR or radar.
So how LiDAR be implemented more cost effectively and power effectively?
A More Efficient Laser Driver
A significant step to reducing the cost and power budget of LiDAR lies in the laser driving circuitry, which is the most power-demanding part of the system.
For LiDAR, the laser driver must deliver extremely high-power, short-duration pulses to the laser diode. To achieve long-range measurements, optical power of 400-500 W is needed, requiring the driver to supply peak power of around 1 kW. For accuracy and eye safety, these pulses must be incredibly brief… typically 2 ns or less.
Figure 3: Schematic of Silanna’s SL2001
Traditional designs use a discrete component approach to meet these demands. A resonant circuit is created to overcome parasitic inductance, but this requires a high voltage (around 100 V) supplied by a boost voltage regulator. This discrete implementation is inefficient, especially the isolation stage between capacitors, which loses more than 50% of its transferred energy. In battery-powered devices, boosting a low voltage requires multiple complex stages, compounding efficiency losses and reducing battery life. This inefficiency generates significant heat, often necessitating large heatsinks that increase the module’s size, complexity, and (crucially) cost.
A more advanced solution is to integrate the boost stage, GaN FET driver, and control logic into a single device. This approach, seen in Silanna’s FirePower™ driver ICs, avoids the primary energy losses of discrete designs. By integrating these functions, it is possible to save 90% of charging power losses and achieve an overall charging efficiency of 85%. These single-chip drivers provide precise pulse control and on-chip fault monitoring for eye safety. This high level of integration enables the creation of smaller, more power-efficient, and cost-effective LiDAR modules for next-generation vehicles and a wider range of handheld applications.
For more information on Silanna’s range of laser drivers for LiDAR and range finders please take a look at the Laser Driver FAQ, or visit the FirePower page.