Blog

One thing most OEMs are doing when they approach designing a new vehicle is they iterate on something that’s already done before. That’s good. (I mean) There’s lots of history there- evolution rather than revolution, but it’s very rare that you get the chance in an automotive career to start on a blank sheet of paper. This was the opportunity and that’s because it doesn’t make sense to take what’s been developed for the last 100 years to solve conventional human driving. It doesn’t make sense to take that platform, and to use that platform for what comes as the third generation of autonomous driving.

Zoox has chosen to develop a vehicle from the ground up for autonomy because we firmly believe that simply putting sensors onto a current passenger car is fundamentally flawed. It’s fundamentally flawed in that the current geometry, shape if you like of a vehicle doesn’t allow you to place sensors in the optimal position so we start sitting back saying “so what do I need to enable autonomy? I need EV, I need optimal sensor positions, I need a large battery to be able to deploy for the day”. We take all those 3 factors, you have to develop a vehicle from the ground up, leapfrogging the rest of the industry, and go straight for that third generation of vehicle, which is EV and autonomous, looking to the future, solving it now rather than iterating on a current platform.

Andy Piper, VP, Vehicle Development

Background

Zoox has always been the stand-out player in the Autonomous Vehicle space, insisting on building a vehicle from the ground up. Perhaps driven by the successes of the Stanford team at the DARPA Grand Challenge, and detracted by the hardware struggles of the CMU team, the route to autonomy has been seen as a software challenge to varying degrees.

In the most extreme case of hardware aversion, the comma two devkit is a USD999 product which crams hardware required for an Advanced Driver Assistance System (ADAS) into a device no larger than a smartphone. A more mainstream example would be Tesla, whose “Full Self Driving” package (which is but an L2 system out of 5 levels on the SAE scale) does not utilize a LiDAR.

3 years later, in 2017, Waymo went back on the conventional route, retrofitting Chrysler Pacificas sparking an outpouring of emotions by automotive journalists, with Quartz proclaiming that “The cutest thing Google has ever made is dead“.

While the Firefly did demonstrate some considerations that an Autonomous Vehicle should be subject to, it was far from radical. In truth, it would probably have been difficult to pull off at the time, given the pioneering nature of the work. In Autonomy, Lawrence Burns (former corporate Vice President of Research and Development at General Motors) plays up the non-technical challenges surrounding AV development at the time, including the vastly contrasting cultures between Silicon Valley and traditional automotive manufacturers in Detroit. I think we can agree that the design of the Firefly can be considered a success, putting an almost cuddly appearance on a piece of technology that few knew anything about and would harbor doubt towards.

At about the same time that Project Chauffeur was announced, Zoox was founded. Understanding that retrofitting is sub-optimal, they looked to develop a platform from the ground up that was symmetrical, bi-directional, and battery-electric to take a crack at autonomy, with the promise of a massive payday. By 2018, they had shown off a functional prototype to Bloomberg which dared to be different, with redundancy in their critical systems and incredible agility thanks to its all-wheel and crab-steer capabilities.

Ask any roboticist and they will undoubtedly tell you that integrating hardware and software is a mammoth task. With just software alone, iteration comes with a cost. Add hardware into the mix and there is an additional layer of balancing (whether one should solve a problem in hardware or software) and exponential time/material costs associated with debugging and iteration, on top of the prices of each item in the final Bill Of Materials. Yet understanding the self-driving problem through the eyes of a roboticist generates huge returns in the long run. With this unveiling, not too far behind the deployment timeline of other projects undertaken by companies such as Waymo and Voyage, Zoox has reaped the dividends of trading off initial development costs (time, financial etc.) for subsequent development rate and scope of possibility, only made possible by a radically different approach which understands and eradicates the constraints of the traditional automobile form.

A quick look at some technicalities

Crab Steer

What it is

A holonomic system is one where all Degrees Of Freedom (DOF) are controllable. For simplicity, a regular automobile can be assumed to live on a 2 dimensional plane. To fully describe the position to such an object, we would need 3 parameters- 2 for position (typically x and y), and 1 for heading (commonly theta, but most letters of the Greek alphabet are cool).

To get an idea of how difficult it is to control an increasing number of DOFs, one simply has to look to drone racing. With 4 control inputs (x, y, z translation, which are coupled with roll and pitch, and yaw rotation), even the most skilled pilots lose control, resulting in crashes all the time.

While it seems that the Crab Steer implemented in the Zoox vehicle is not fully holonomic (partially due to geometric constraints imposed by mechanical linkages used for steering), it allows for the car to maintain its heading while translating in the x or y axes within a given geometric bound. This is demonstrated when the vehicle changes lanes in the Bloomberg video. If done in the absence of autonomy, a human driver would require an additional control input, like a second steering wheel, which is a terrifying thought considering 1.35 million people are killed on the roads each year even with a relatively simple control scheme.

Why it’s useful

If human drivers have been doing just fine with 2 control inputs, why would we need more DOFs? In general, it expands the action space, which fundamentally affects the way we drive, and the way our road infrastructure is designed. For the less spatially-inclined, we will now be able to parallel park with virtually no gaps between neighboring vehicles, greatly increasing parking density (a major pain point in the design of urban cities). The impact on safety will also be profound, as Zoox vehicles will be able to translate laterally in an almost instantaneous manner to avoid crashes, or even perform complex maneuvers, while drivers in traditional vehicles can only slam on the brakes and wish for a favorable outcome. During the unveiling, the Zoox team also brought up points pertaining to efficiency of movement (turning with a much tighter radius, no need for 3-point or U-turns, etc.), which would perhaps appeal more to the average consumer.

Protruding Wheels

The first thing that struck me about the Zoox vehicle were the protruding wheels, simply because they have been a primary consideration in my designs for Autonomous Ground Vehicles.

The rationale given by Zoox for this design choice was simply that it allows for a more compact vehicle. However, it is so much more than that.

In conventional cars, especially lowered ones, drivers may find that the front bumper could come into contact with raised curbs and speed bumps due to limited ground clearance. However, this is not a needless inconvenience. Having wheels tucked into the chassis, I believe, is done in part to accommodate the space occupied by the engine in traditional cars (other reasons could include the poor aerodynamics of a box). With a radically different layout, as mentioned by Zoox, electric vehicles which still have their chassis protrude beyond the wheel are holding on to an irrelevant artifact of gas engines. They needlessly put an expensive sensor stack right in harms way in the event of a crash, and results in a major safety concern akin to strapping the driver in front of your car.

With protruding wheels, the Zoox vehicle can not only scale uneven terrain with ease (as repeatedly shown in the unveiling video), but also ensures that the sensor stack is not the first point of impact in a collision. Other advantages may include a reduction or elimination of the dead zone (due to minimum range) of rangefinders due to its recessed position within the overall footprint of the vehicle, and slightly better stability due to the concentration of mass within the base area.

Final Thoughts

Is the design of the Zoox vehicle perfect? Most definitely not. A common gripe I have would be the positioning of LiDARs on the corners of the vehicle. While great for maximizing all round visibility, it leaves a critical sensing component exposed and vulnerable to impacts. The cantilevered design also raises questions about the rigidity of these mounts, which impacts the accuracy of geometric transforms used when fusing data from multiple sensors (that is, until we can figure out continuous robust online calibration).

Will Zoox emerge as a frontrunner in the AV space? It is difficult to speak with any level of confidence. Bringing a product to market requires much more than a smattering of good design considerations from internal factors like finances, deployment strategies and the actual ability of the driver to external ones like public perception and legal hurdles.

That said, Zoox has managed to concoct an aesthetic that is friendly, elegant, and oozing with brand consistency, while ensuring that functionality remains a first-class citizen. It has shown that there are benefits to designing a vehicle from the ground up, which can be obtained in a reasonable timeframe (as long as you have Amazon swooping in to save you, of course).

As a roboticist, I am proud of how Zoox is representing the industry and showing that there is merit in investing in hardware development through a robotics approach in a world overly enamored by software as a silver bullet.

I’m looking forward to so much more.