Zoox and Its $800M Robo Taxi Dream [video]
bloomberg.com
bloomberg.com
https://www.dmv.ca.gov/portal/wcm/connect/495c1aa7-2a3b-4dd8...
For comparison, in the 12 months ending in Nov 2017, Waymo drove 352000 autonomous miles in CA public roads, with a disengagement rate of 1/5600 miles.
This article is a joke...
https://thelastdriverlicenseholder.com/2018/02/01/disengagem...
They seemed pretty confident when they had a reporter ferried back and forth.
It's not bad being in 3rd place against Google and Cruise/GM in a few years.
This type of thing can improve rapidly as they start doing real-world testing... which you cite yourself has merely begun (on real roads mind you, not testing in closed environments). 2200 miles is hardly anything to measure against. Which sounds like months of testing vs Waymo's years.
Not to mention they are developing full new vehicles as well, not just the driving AI software. Which might be highly valuable to a major car company who could acquire them at a significant multiple.
This type of pessimism based on a single metric is toxic and naive regarding how tech companies are built.
...Not to mention there was no claim in this article that they were at Waymo's level or anything but at a very early stage of development.
[1] https://www.crunchbase.com/organization/zoox#section-locked-...
The self-reported California metrics themselves cannot be relied upon since the companies (Cruise and Waymo included), could be making it up (and indeed have an incentive to inflate their numbers).
At the end of the day though, not all miles are equal. If one were so inclined you could just drive up an down a big highway with autonomy engaged only once merged. Obviously 100 miles of SF traffic is not the same as 100 miles of staying in your lane on cruise control.
> This seems to violate
> the classic startup maxim
> to 'focus on just one
> thing and get it right'
how many companies that you know of that are working on self-driving cars have done this? had a reporter do a similar route in a car driving itself? I know nothing of this company but they are either completely insane or have a high degree of confidence in their product, or the reporter is lying. I have no reason to suspect the reporter of lying in this case.
What was wrong with the turn? They had to wait until they could confirm that the second oncoming car would actually slow and complete their turn before it was safe to continue. That's how many humans would handle that situation too IMO.
also, why does it have to be vision only and no maps? the goal is not some pissing contest of who has the cleverest algo, the goal is a working product. Of course, it's going to include maps.
The reason I'd want to see a vision-only mappless demo is because Lidar and HD maps are a crutch, which make it easy to create a 90% functional demo, yet are insufficient to cover the long tail end, which is inherently a computer vision problem. There is not enough information density or semantics in Lidar data to make high level judgements, which makes it useless for many relevant corner cases. Meanwhile, vision has all the information but it's a hard yet necessary AI problem to extract it.
The goal is a working product indeed, but if it involves teleoperation, human annotation for maps, and over-relies on Lidar, it will be a slow and unreliable AI granny driver, which will penetrate at most 1% of the market. There is a whole lot of unscalable human effort being swept under the rug in most autonomous demos, so don't believe everything you read in the media.
It's frankly sloppy that the reporter doesn't mention it. He also incorrectly reports that Waymo only tests in places like Arizona.
"Overall, the vehicle performs so well that you forget no one is driving."
Jesse Levinson, the guy leading autonomy at Zoox is a big scoop, he gets huge respect in the robotics community. When you got a top talent like that, it tends to attract more top talent. I'm really impressed by how much Zoox has accomplished with so little. Last I heard they had about a dozen test vehicles noodling around parts of San Francisco, but with this new cash injection you'll probably be seeing a lot more sensor riddled flat black Toyota Highlanders in the Bay Area in the coming months.
It should be emphasized that fully validating an autonomous vehicle well enough to remove the driver is no small feat and they have a long way to go, I'm betting against their 2020 target for a commercial rollout.
I'm somewhat baffled by how much spite a tech news site like HN has for technology companies that are actually innovating and solving hard problems. Zoox didn't drop out the sky yesterday, they've been at this for years, you just haven't been paying attention.
He said "no one is driving" Now, maybe he is full of shit, but those words are pretty definitive.
This reminds me so much of dot com boom bravado and then hubris. If these guys had built lots of different types of cars from the ground up - race cars, customs, hot rods etc powered by different engine types - I would have more time for their reinvention confidence. I like what they are doing but this seems like pride before the fall
"If we are right, then we're going to dominate the market that all the big boys are investing billions to control."
(Not being disparaging at all - their value proposition actually looks interesting.)
The line about disrupting a huge market is still par the course, what's expected of them.
Zoox is the only company we know of that is building a robotaxi from the ground up rather than just retrofitting a vehicle designed for something else. Although Einride has a concept built for a highway freight hauler with no driver cabin, and Nuro has several delivery vehicles built with no room for a human on board.
The term 'self driving car' is kind of like the term 'Horseless Carriage'. It's transitional.
Give that, I'll say this. Mobile robot from the ground up = next level hard. Level 5 autonomy = next level hard. Both simultaneously? Good luck.
Question though, what if there's a regulatory requirement for future vehicles to have a mandatory human driver mode (manual override) that's a touch of a button away?
Cruise Automation was going to do exactly the same things as Zoox. Cruise sold for around $1Bn to GM. Cruise had 40 employees and had basically sold nothing.
How about developing the world's leading level 5 system and then licensing it to the all auto companies?
There is another interesting, but still young company called Aurora innovation that has a very good team together that plans to develop and licence L4 to 3rd parties. Aurora has partnerships with 2 of the top 3 biggest car manufacturers in the world, VW and Hyundai, as well as with Byton, a Chinese EV startup backed by big Asian capital. Zoox has a different plan.
(edit: changed link to point to wikipedia instead of random site)
A whole essence of investments fraud, dragged to the edge of the sense. And they politely call it ‘pitch’ in this thread. So cute.
Let's say I'm an investor and a founder of a company in my portfolio does this. What would be an appropriate response? This seems like a cartoonishly gratuitous waste of capital.
You'd either fire them on the spot. Or, if that's not an option [1], quietly figure out a longer-term strategy for eventually firing them.
Or, if you're a Nice Guy, you'd offer to let them stay on if they return the gratuitous waste of capital and/or pay the company back.
I'll never quite understand why founders are treated as an exceptional case by capital. Especially in hard tech fields like self-driving where the idea is literally obvious to a five-year-old but the execution requires courting `the best humans available.
--
[1] e.g., an engineer with mostly undocumented knowledge of key parts of your software or sales rep with close relationships to a client during a critical time period.
I have seen termsheets where the founders need board/investor approval for spend over $XXXK, so that it stops/limits misuse of capital.
This could have been board approved, if it needed to be or else investors/boards can always ask for it to be returned and mandate approval for future spends above $X in categories outside of regular business spending.
Investing in someone with zero useful ideas and zero ethics is not a symptom of good investment skills. (Yes, tricking experts into talking to you is unethical)
I half-suspected that they didn't bother taking out a permit for autonomous vehicle, but I was pleasantly surprised to find them on the list: https://www.dmv.ca.gov/portal/dmv/detail/vr/autonomous/permi...
That's what Uber considered over-regulation.
Building a little vehicle for local transport is reasonable. At least five other companies have built minibuses for parking lot shuttles and such.[1] Nobody seems to have more than pilot installations. This is embarrassing. Under 25MPH self driving ought to be within current technology. At low speeds, you can stop fast enough that predicting the behavior of others isn't necessary. Navya has a few installed systems, but they're either on dedicated roads or under 10mph.
The fact hat it's not tells you all you need to know about self driving vehicles. I'd expect full autonomous vehicles about a decade after we can hit these easier targets.
The "fake it til you make it" startups aren't succeeding at this.
If you could build the pods on the cheap, cities would love it as an affordable alternative to subway and tunnels - it's much cheaper to dig a 4x2m concrete box under the main road then massive tunnels and stations.
In time, as the network expands, full self driving can be added and the vehicles can exit the network and drive the last miles on regular roads to the traveler's front door in the suburbs.
It still costs money. And where should the exits be? If they are far apart this doesn't work. And the prospect of being underground for a long trip is not that thrilling.
You're basically proposing a boiling the ocean kind of solution.
When passengers are secure, another ramp takes them back down - or even the same one, with proper synchronization on less busy stops. It would then drive the passengers to any station in the network with almost the same flexibility as a self driving car.
I would say the cost differentials to subway are massive, one order of magnitude for similar passenger throughputs. And it's not some unproven, highly speculative technology like self driving cars, it's something that could be built today with low tech (but I concede, high capital investment).
Self driving pods could work but you would basically need to dig tunnels as big as current subway tunnels to deal with the traffic + breakdowns etc.
Bear in mind, a subway tunnel for a single train has a diameter of about 7m, or about ten times the area (and dirt volume to be moved, and concrete to be poured) of a 2x2m lane.
>You should also factor the much higher efficiency of a system with no train changes, where the pod occupancy is optimized to avoid those empty returns to the busy area
How do you avoid empty returns to the busy area? You still need to supply the pods, and if your solution is underfilling pods on the way back, that's exactly the same as sending a light train.
>where you don't have to dig a whole tunnel worth of capacity (6 tracks) at the terminus stations since one is enough for the traffic there
This is only an advantage at small scales. At higher scales, one train tunnel is just enough at the terminus stations, and the busier routes have more.
A 2 second gap means that a pod needs to sync and squeeze between two vehicles 4 seconds apart, or 45m at 40Km/h. That's three seconds of 0.3g on a 40m acceleration track, after which it can merge with no assistance from the other pods (in a practical system, they would clearly cooperate). Assuming you achieve 90% of full 2s capacity, a pod will wait an average of 20 seconds for it's slot - queues should not form outside of very busy stations, that would be connected to very thick pipes anyway.
> How do you avoid empty returns to the busy area?
You might need to return empty, you might not. There might be a triangular route available that can maintain fullness throughout. You could build a buffer in anticipation of the rush hour. The train absolutely needs to follow the same route back at the same time interval and there is no way to reallocate part of the return capacity to the forward, overloaded route by a simple lane direction reversal. (to clarify, I am talking about maximizing peak capacity at rush hour on a given overloaded segment, there would be significant empty runs during the rest of day like in any other system).
> At higher scales, one train tunnel is just enough at the terminus stations
That's because you only have underground trains to places where a full tunnel makes economic sense, so the city is substantially under-served.
Sorry, I don't follow. Why can't a 2 person pod reach 40 Km/h and maintain 22m spacing between pods?
As for the number/price of the pods, the relevant metric is price per passenger. A Subway train costs about $10.000 per standing seat, with the bulk of costs going to infrastructure, not rolling stock. So with cheaper infrastructure and a $20.000 2 person pod, we are cost competitive with standing seats in a packed subway, at much higher passenger comfort.
2 seconds at 40 Km/h is safer but that has other problems.
Remember, Pods don't have zero length they take up space which decreases the pods per hour. A 2 person 3 meter pod with 22 meter spacing at 40 Km/h = 2 * 40,000m / (3m + 22m) = 3,200 people per hour if they are all 100% full.
That's a long way from a subway's capacity unless you want to build 20 lanes per subway you are replacing. Going faster means more pods, but you also need to increase following distance. Slower is safer, but means trips take longer.
Everything that is expensive and unreliable is because they assume an embedded wire is not going to be allowed/done by the people managing the roads.
But it’s NC State and not Top4, so don’t expect anything to come out of it.
All while avoiding the seemingly insurmountable difficulties of fully autonomous driving on congested and chaotic city roads.
For example, I worked with a company named "Mobility Engineering and Development" outfitting Ford vans that C-4, C-5 quadriplegics could drive while sitting in their wheelchairs.
Ford didn't redesign those very often so it was a very good choice for that purpose, but it would be a nightmare to have to redesign what we were making every 3 years to accommodate a new vehicle design produced by a 3rd party that had no interest in what we were doing with them.
If the reports that Apple is building a similar product are true, there's a hell of a conflict of interest here.
This... hubris hurts when it's not in the show Silicon Valley. At least the show hurts good.
Cause if I can pay a highly reliable driver for less, what's the point?
Right now they’re fighting to own the platform that works when, ten years from now, the technology is affordable. That means developing on higher priced technology now to get the software sorted out.
Human drivers that are cheap probably leave a lot to be desired in the “reliable” category (I conjecture), so robots really can (in theory) offer more there.
So you easily get to 5x a good driver's annual cost probably being cheaper.
Let's assume that the technology and mapping becomes a Solved Problem.
Then the competition with humans boils down to the cost of sensors and compute (plus recouping R&D). I expect prices for both of those to plummet whenever self-driving capabilities are clearly demonstrated because the massive anticipated demand will justify huge capex in improving efficient manufacturing of these often bespoke sensors.
The market for computing machines 1950 - 1990 is instructive.
Zoox's initial fleet will be hand built, and will likely cost several hundred thousand dollars a piece to manufacture, but hypothetically one could still turn a profit on a robotaxi worth $300k.
Bullish analysts have speculated that economy class robotaxis will someday be cost competitive with public transit, though I suspect it will take a while to get there.
Of all the robotaxi companies out there showing real progress, Zoox is the most differentiated, and is positioned well to avoid the race to the bottom by capturing the premium end of the market, much like Apple has done with smartphones.
Of course this is all contingent on Zoox's ability to execute. But even if they fail miserably at developing their bidirectional robotaxi with independently articulated steering and drivetrain for each wheel, their autonomous OS is still very good. If we compare Zoox to what other companies such as Apple, Intel, Baidu and Uber have spent to accomplish less, Zoox's autonomous OS alone is arguably worth the $3.2 billion that the company is being valued at. So it's not like they don't have a plan B.
Yeah, everyone who has putzed with robots gets a tingle of excitement when they build their first crab drive. It's neat, but also introduces a huge amount of complexity and probably maintenance wear on the tires from non-subdegree alignment at road speeds.
Tesla is struggling to mass produce electric-from-the-ground-up traditional front-wheel-steering-only powertrains. What's the use case for this degree of added complexity beyond some cute parallel parking?
Otherwise you're right, it's very much an open question as to whether they can make these radically different vehicles to automotive grade reliability.
"...It’s in the city, though, where Zoox really shines. The screens inside the vehicle show an overwhelming amount of information, as the computer vision software keeps tracks of cars, people, stoplights, and road markers all at the same time. Unlike many self-driving cars, it glides to stops. At an intersection with a left turn, it allows oncoming traffic to pass and then waits for some slow pedestrians. Overall, the vehicle performs so well that you forget no one is driving. ....."
?
How would the reporter know if the car was or was not being remote control operated by an employee?
Is “overwhelming amount of technical info” any more impressive than those memes of computer hackers typing really fast with the green terminal text scrolling down like you see in the movies?
The left turn doesn’t sound like anything an autonomous car shouldn’t be doing? He’s wowed that the car stopped for a pedestrian? Really?
Coincidently Ericson and KTH (which Zoox has cooperated with) has done remote driving over cellphone links over multiple miles. https://m.youtube.com/watch?v=cvSM7pbqZEM
I'm not saying it's what happend. I mean Darpa challange was not fake so Stanford has delivered in the area. But it's possible to solve the lag issue.
But of course we know Zoox is actually driving on SF roads according to public documents, and the fact you can watch them drive if you are in SF and know where to look. Demos are demos, but this one happened with Zoox software at 100% autonomy.