Today, Tesla, Inc. 's second AI Day is finally officially held.

AI Day is one of Tesla, Inc. 's annual activities. Other regular activities include Battery Day and Autonomy Day, which, as the name implies, are specifically for AI, battery technology and self-driving related activities.

This year's AI Day has been delayed for more than a month. Musk revealed last month that the purpose of the jump was to allow more time so that the available humanoid robot "Optimus" could be displayed on AI Day.

Sure enough, Musk didn't let everyone down:

We saw Optimus Prime's prototype product (code name seems to beBumbleseed） ：

In order to attract attention, Musk even announced in advance that there must be a "Catwoman" version of the sky pillar robot in the future.

It has been developed for only 8 months and has the ability to act independently

Despite the gimmicks, let's take a look at how powerful the prototype robot is.

The prototype robot, codenamed Bumbleseed, does not have the shell we saw on the Tesla Bot mannequin last year, and all the internal non-critical parts are exposed:

The prototype robot stands about 5 feet 8 inches (about 173cm) and weighs 73 kilograms.

It basically uses the same equipment of Tesla, Inc. electric car in the part of perception and computing, including the chip integrated system independently developed by Tesla, Inc. as the "brain" and three Autopilot-level cameras.

In addition, the head of the Tesla Bot designed last year also has a display screen that can display useful information for users / partners. There is no progress in this design that can be announced today. This is because the team has focused on solving locomotion rather than application problems over the past year.

The prototype robot's walking speed is still very slow, which seems to be far short of the 5 mph (about 8km/h) level claimed last year.

As for carrying capacity, it can carry about 20 pounds (about 10kg) of small and medium-sized goods, such as express cartons. This load is half less than that announced last year.

Optimus Prime has 28 motor "joints", down 1/4 from 40 announced last year. This is because the fewer motors, the lower the power consumption.

In the sitting and standby state, the whole system power consumption of the robot is only 100W, which is less than half of that of the mainstream graphics card, while in the walking state, the whole power consumption is only 500W.

Musk and members of the robot team have repeatedly stressed that making robots save electricity is a very important design concept, because Tesla, Inc. 's ultimate goal is to produce humanoid robots on a large scale and put them into use in a variety of industrial and civil scenes. The order of magnitude is in the order of millions or even tens of millions.

The blue one in the picture below is the electronic system. The orange part is the motor.

Let's first take a look at the electronic system part:

The robot is designed with a battery capacity of 2.3 kilowatt hours and a voltage of 52V; the "brain" part of the robot is not really placed in the head, but installed in the torso position (because it is safer here, and the more integrated it is, the more power is saved).

The robot supports Wi-Fi and LTE in terms of network connectivity and includes audio components, all of which are installed on the torso.

Let's take a look at the machine's athletic ability.

Members of Tesla, Inc. 's robot team revealed that they used the same prediction model as the electric vehicle crash safety test to predict in advance the performance of the robot in various positions when performing normal work, and to predict a variety of situations that could lead to damage to the robot, such as falls and attacks.

Then, the engineering team carried out a large number of simulations on the collected data of various working conditions by means of calculus, so as to fine-tune the working power of the motor to minimize energy consumption; in addition, they also found commonness from different motor data, thus reducing the use of different types of motors.

In the end, of the 28 motors in the whole body, there were only 6 different designs, which significantly reduced the cost of robot mass production.

Among these motors, there are large motors with strong load capacity, which are large enough to easily lift the weight of a grand piano, and small motors with small but sensitive hands that can not only grasp goods with a certain weight. it can also accurately grasp small parts, operate more sophisticated tools and instruments, and so on.

Of course, in addition to using mathematics to solve problems, designers have also added some bionic designs. For example, the knee joint of a robot is divided into four sub-motors, which is similar to that of human knee joints / ligaments.

Musk said last year that the mission of humanoid robots is to replace humans to do dangerous, repetitive and boring work.

Its initial idea is to allow users to give simple commands to it, such as "install the light bulb", and it can be executed.The company's positioning for Optimus Prime is a "universal AI robot", which is, to put it bluntly, a coolie robot. Man-machine intelligent communication is not one of its design capabilities.

Although Tesla, Inc. has all kinds of problems, it is irrefutable that the company has come to where it is today because of its crazy pursuit of scientific and technological innovation and Musk's own strong will that no one can stop. Last year, many people wondered whether Tesla, Inc. really wanted to build a robot or joked with everyone. As a result, Musk really kept his promise this year.

Maybe one day Optimus Prime can really replace Tesla, Inc. 's workers in car factories around the world-their jobs are really dangerous, repetitive and boring.

Musk also said last year that the robot would not run too fast. "our design indicator is that you can easily not be overtaken by it and have a good chance of subduing it."

"it's best that this doesn't happen, but who knows." He added.

Since only one dummy can be displayed last year, to this year there is a prototype that can run independently. Tesla, Inc. 's humanoid robot team has made great progress in a very short period of 6-8 months, and he is very satisfied with the team's results.

In the coming weeks to months, Tesla, Inc. will continue to assemble more robots and test them in real industrial scenes, such as Tesla, Inc. 's own Fremont factory in California, to further optimize the design.

The exciting thing is:The robot has conceived to have produced a prototype that can move freely without any external support (power cord, protective frame, etc.).

The importance of this situation is that although today's prototype robot is still a long way from Boston Power's Atlas in technology, Tesla, Inc. 's development speed is significantly faster.

Last year, Musk said on AI Day, "Tesla, Inc. is already one of the largest robot companies in the world." Our cars are actually intelligent robots with wheels. "

Through the Optimus Prime project, Tesla, Inc. finally transferred the company's years of research and development in robotics, sensors, AI and other technologies from four wheels to a humanoid robot with two arms and two legs.

A member of the team said"We are confident that we can achieve mass production of humanoid robots in the next few years."

Musk pointed out that humanoid robots on the market generally do not have intelligent brains, and the cost is too high and the output is too low.

Tesla, Inc. 's long-term goal is to reduce the production cost of humanoid robots to about $20, 000-less than the price of a Tesla, Inc. electric car.

"We still have a long way to go. So today we are holding this event in the hope that smarter people will join us to make available reliable humanoid robots a reality and large enough to help hundreds or even tens of millions of people. "

Dojo supercomputer

From 2019, Tesla, Inc. launched a mysterious project, Project Dojo, to develop a multi-chip modular (Multi-Chip Modularized) supercomputer with self-developed chips that can be seamlessly connected to achieve ultra-large-scale distributed computing.

This supercomputer is the Dojo that was shown on AI Day last year.

The core of the Dojo is a D1 processor independently developed by Tesla, Inc., which uses Taiwan Semiconductor Manufacturing Co Ltd 7-nanometer process technology. the chip contains 50 billion transistors and 11 miles of line length. The processor uses Tesla, Inc. 's self-defined instruction set, the FP32 computing power reaches 22.6TFlops, and the design power is about 400W.

However, computing power is not that important compared to the processor's Ithumb O bandwidth. This is because the main limitation of the current self-driving computing platform is not computing power, but data transmission, and this supercomputer is trying to solve this problem and speed up the rapid perception, calculation and data transmission of the real world. as a result, the training performance of the autopilot model is significantly improved.

Last year, Tesla, Inc. claimed that the D1 processor used by Dojo has much higher off-chip bandwidth than the GPU-based distributed computing array of Alphabet Inc-CL C TPU and other technology companies.

Ultra-high off-chip bandwidth is particularly critical, because for Tesla, Inc., D1 Chip is not used one by one, but can be used to form a computing array. For example, 25 pieces form a training module (Training Tile), which becomes the core of the Dojo unit, up to 36TB/s bandwidth and up to 9PFlops computing power.

Last year, Tesla, Inc. revealed that he had completed the working test of the training module and successfully ran the minGPT model developed by Andrej Kaparthy, the former director of Tesla, Inc. 's AI department. In the past year, Tesla, Inc. continued to optimize the research and development of Dojo.

At today's AI Day, the Dojo team claims that the machine learning training power of one training module is enough to reach six "GPU computing boxes" and that the cost is less than "one box".

This is not over: six training modules are combined into one layer, and the two layers are stacked together to form a training matrix (single cabinet).

Then link the cabinet to form a "complete", that is, Dojo computing cluster ExaPod, including 120 training modules, 3000 D1 processors, FP16 computing power as high as 1.1 EFlops.

Prior to this, the Dojo system mainly served the internal use of Tesla, Inc., such as road test data labeling, model training, running simulation and hardware-in-the-loop tasks.

But more importantly, Dojo is essentially a computing device with ASIC attributes, which can be regarded as a general neural network training equipment, and can also be provided to other institutions.

Maybe in the near future, we can see that Tesla, Inc. is not only an electric car / clean energy company, but also a robot company, a chip company, or even …... Cloud computing company?