Tesla unveils new D1 chip and Dojo supercomputer to train, improve Autopilot – humanoid robot previewed – paultan.org
Tesla recently concluded its AI Day 2021 event, where it revealed quite a few important things in relation to, well, artificial intelligence. No new vehicles were presented during the event, but we do get a preview of the Tesla Bot, which is a humanoid robot that the company actually plans to build, but we’ll get to that later.
Currently, all new Tesla models are equipped with the Full-Self Driving (FSD) computer that provides the processing power needed to enable Autopilot features. This is linked to the radar-less and camera-based Tesla Vision hardware package that captures information about the car’s surrounding, which is then sent to the all-important AI software that decides how the car should approach a driving situation without input from the driver.
Despite the marketing name, FSD is only rated at SAE Level 2 automation, so drivers will still need to be at the ready to take over when the system is unable to cope. Of course, Tesla is working hard to achieve the coveted Level 5, where a car can drive itself under all conditions without requiring the driver to step in at all.
It’s certainly not an easy goal to reach because traffic behaviour is extremely complicated, more so for an AI that must be trained to deal with a variety of situations. As such, a lot of data is required with different driving scenarios for an AI to learn, which isn’t much of a problem for Tesla as its customer fleet on the road today can log millions of kilometres of driving, with or without Autopilot active.
When an AI software meets specific standards or requirements, only then will it be rolled out to customer cars via over-the-air updates. With new software working with the hardware, more features and functions can be made available to owners, which can enhance the capabilities of Autopilot and FSD as a whole over time
However, before anything gets sent out, Tesla must ensure that it has accounted for all possibilities based on real-world conditions, which require AI models to be trained based on data it has gathered. For this, the company unveiled its new D1 chip at AI Day, which will be used in its new Dojo supercomputer system to continue its pursuit of vertical integration.
Built using a seven nm manufacturing process, the chip boasts 362 teraflops of processing power with 50 billion transistors, which is plenty of oomph to crunch through data. The company says 25 chips are used to make up a single “training tile,” and 120 of these tiles will be packaged together across several server cabinets for over an exaflop of power.
According to Tesla CEO Elon Musk, Dojo should be operational by next year, and that the reason for taking chip production in-house is to increase bandwidth and decrease latencies for better AI performance. By owning as much of its tech stack as possible, the company can also reduce the possibility of bottlenecks.
All the processing power and benefits will be used to train AI models for recognizing a variety of items from video feeds collected by cameras on Tesla vehicles, which trains the AI to be more effective and with more capabilities. The company says these strategies helped it retire the radar sensor from its cars, and there’s even a demonstration of how the tech has improved its Autopilot system.
Musk notes that its supercomputer tech can be used for many other things and that Tesla is willing to open up other automakers and tech companies who are interested. This brings us to the Tesla Bot, which is meant to showcase how the company’s advanced technologies can be used in other applications outside of cars.
There’s no working prototype yet (Musk says it could arrive next year), so a Tesla employee dancing in a suit on stage is all we get for now, but the renderings do conjure up scenes from Will Smith’s 2004 film I, Robot, which has both AI and human-like robots.
“Basically, if you think about what we’re doing right now with cars, Tesla is arguably the world’s biggest robotics company because our cars are like semi-sentient robots on wheels,” Musk said at the event. Basically, it’s all of Tesla’s developments packed into a humanoid form, but with the aim of “eliminating dangerous, repetitive and boring tasks.”
First thing you need to know about the Tesla Bot is that it will be friendly, just like Honda’s Asimo was, so you need not worry about it going on a murderous rampage, unlike in I, Robot. Secondly, despite its height of nearly 173 cm and 40 electromechanical actuators, it isn’t particularly fast – the top speed is just eight km/h. “We’re setting it such that it is, at a mechanical level, at a physical level, you can run away from it, and most likely overpower it,” Musk joked.
With AI Day over, it’s clear that Tesla has its own vision when it comes to autonomous technology and is approaching it on its own terms. With the D1 and Dojo, it can now accelerate the development of its FSD and Autopilot systems, the former of which is still in beta phase, bringing it closer to offering Level 5 autonomous cars. As for the Tesla Bot, let’s hope that the AI keeps to Isaac Asimov’s three laws.
Source: Read Full Article