Have you seen a RoboTaxi? I live in Santa Monica. Two robots roam my neighborhood streets freely these days (c. 2024), and one waits in the wings. There are Waymos (waymo, the robotaxi JV between Google and GM), there are Cocos (Coco Robotics, the ultracute food delivery robots), and there is… Tesla Optimus (the humanoid / human replacement robot… for what?). Oh! And there are also, many many… Teslas, driven by the v12 AI FSD.
Why Tesla Wins: It’s All About the Training Data, Stupid
Everybody else is putting $300k of hardware, primarily in the form of LIDAR and GPUs, atop existing commercial vehicle platforms (Waymo uses a $100k Jaguar ____… Nuro uses $25k Toyota Priuses) in order to attempt full selfdriving competency.
Everybody else (Uber, Waymo, Toyota) is using human-coded algorithms, based on 50+ years of computer science theory, to drive their AVs (autonomous vehicles). Tesla, too, followed this path, for more than a decade, pursuing the Holy Grail of FSD (Full Service Driving). More amazingly, in these days of “pre-order” KickStarters and AI vaporware, Tesla began charging a subscription to FSD in 2016… long before the tech was “ready for prime time.” This option can either be bundled at point-of-sale for the paltry sum of $15,000 (yes, Fifteen Thousand Dollars, USD. You read that right. About the cost of a well-maintained, pre-owned 2016 BMW 5-series car).
Alternatively, for those who might shy away from a not-quite-functional $15k vanity purchase, in July 2021, Tesla offered it as a monthly subscription option. For those who value their commute time and would rather be checking e-mail than steering and braking, for the modest sum of $199 a month (hey, that’s still less than an Equinox membership!), they too could experience the risky rapture of (mostly) hands-free FSD.
The “flaw” (imho) of Google and GM and every other wanna-be king-of-the-AV-road is that they all depend heavily on LIDAR (roof-mounted 3d laser scanners). LIDAR produces a massive flood of real-time 3D data that requires an 800w supercomputer in the trunk to process (don’t worry, it runs on a separate battery / power system). This is where Tesla wins. if ChatGPT proved anything, it was that “the best training data wins”.
Every Tesla that’s come off the assembly line since 2016 has had 8 or more cameras, streaming HD video 24/7 up to Tesla FSD’s datacenter. When you buy a Tesla, your EULA (required for purchase and operations of vehicle) states “all video data is owned by Tesla, for autonomous driving training purposes.” (it’s true. see: “Who really owns your Tesla? Who really owns your iPhone?”)
Every *day*, Tesla accumulates an additional one *million* miles of footage, which they retain exclusive rights to. Everybody else is left in the dust. The corp with the most data wins.
- link: Figure 01 robots driven by ChatGPT
…why deterministic human-coded algos are dead
. - link: Computer Vision, solved
…why deterministic human-coded algos are dead
“RobotTaxis are our Future” — Musk, July 2024
The central idea is threefold:
- Tesla will be the first to crack the AV (Autonomous Vehicle) nut… the first to truly “solve” the holy grail problem of driving anywhere, anytime, generally without having to painstakingly, manually, map and monitor every inch of a city / driving zone
- Tesla will launch a private fleet of RoboTaxis that compete globally with Uber / Lyft.
- (this is the real innovation imho)
Telsa owners will be able to essentially AirBnB their cars, like a timeshare, so that while they are at work, and while they sleep, instead of their cars sitting idly in the parking lot or garage, they will be in fully autonomous mode, zipping around the city, giving people roborides, and earning money for the car owners.
This last point (provided that the first point is accurate) sounds both compelling and doable to me. It also makes point 2 moot. Tesla ships close to 2 million vehicles a year globally; the pre-existing global fleet already sits close to 6 million vehicles. This dwarfs the fleet of any other competitor. Even if only 1 in 10 owners opted into the program, that would still be a standing fleet of roughly 1 million vehicles in 2026.
Just like the race to AGI, the quest for fully autonomous vehicles is a very expensive R&D pursuit. Far more than $400 billion has been spent across the years, by all the major players combined. Apple famously invested more than $100 billion across 10 years to develop its own vehicle, finally dropping out of the race in 2023 without ever having even demonstrated a concept car.
Other notable AV players include:
We should also keep an eye on robotic delivery services, who are attacking the same problem, with a different value proposition. DeliveryBots have the added advantage that they can use alleys, sidewalks, and park paths. In other words, they can go anywhere a human, or a bicycle, could go.
The most prominent of these is Nuro, which has raised more than $2 billion to date.
Other players to watch include CocoDelivery and Starship Technologies.
Purported rationale for humanoid robots
(for the purposes of our discussion, “humanoid robots” will mean robots built to mimic the general form factor of a human being: two legs, two arms, two hands with five fingers each, a sensor-dense “head”, 150 pounds, 5’10” tall)
bottom line: the entire “built environment” of the world — every building, every bicycle, every sidewalk, every doorway, every car, every ladder, every tool — was designed, either unconsciously or intentionally, for use by human beings — bipedal bodies between four and seven feet tall, with two legs and two arms and a head. with five fingers per hand and opposable thumbs. So the easiest bridge to a robotic future is to design robots to conform to that precise form factor… by doing that, the robots can automatically perform any physical task, in any environment, that humans currently perform… without modification to the environment!
There will obviously be purpose-built robots — one only need visit an automotive or chip-making factory to witness these monsters — but to bridge the gap — and to get robots out of the factory and into the real world — it is imperative that, at least in the short term, that they can operate any tool or interface that a human hand can, and that they can fit into, or move through, or sit in, any space that a normal human would be capable of.
Voila.
related:
- Embodied AI 101
- The Robots are Coming; What’s the Plan?
- Roberts Five Laws of Robotics