Tesla AI Day 2022 officially debuted today. Those who want to watch the fun may be a little disappointed, but for those who want to deep-dive to AI technology, this is definitely a feast of knowledge of artificial intelligence.
Tesla AI Day 2022, as previously announced, will focus on robot Optimus and autonomous driving technology. However, this is not a “product launch conference”, but a “technology exhibition” and “talent recruitment conference”.
“The goal of this conference is to attract more talented people to join Tesla, create new things with us, and change the world.”
Tesla CEO Musk started with the introduction of the robot Optimus, showing Tesla of artificial intelligence technology.
Watch the AI Day Livestream
Optimus Humanoid Robot
The Tesla robot Optimus looks clumsy and rough compared to the well-known Boston Dynamic robot before.
But Musk believes there are two most important keys to humanoid robots that other manufacturers have overlooked: “manufacturing costs” and “a brain that can interact with the real world.”
“We expect the price of Optimus to be mass-produced, which is about $20,000, which is cheaper than our car.”
Musk pointed out that Optimus has a clear production goal, which is to become the productivity of Tesla factories.
Therefore, it is necessary to mass-produce and control costs. In order to achieve this goal, they have imported a lot of car-making experience and applied them to robots.
Many people mocked Tesla for just putting the computer on four wheels as a car, and Tesla in turn quoted the joke, “Now we just put the computer on two legs.” The robot seems to carry a graphics card (GPU) on its torso which some viewer are jokingly mentioned of GPU price rise again.
In fact, Optimus has 28 joint actuators throughout the body, enabling the whole body to produce more than 200 different angles of motion,
The most important hand, with 6 actuators, can make 27 angles, enough to carry a load of 10 kg, and accurately grasp small parts.
In the belly torso of the Optimus, there is a 2.3 kWh battery, enough to provide it with the power it needs to work all day,
Next to the battery is its main computer, the same integrated circuit board as the Tesla vehicle, which is responsible for all operations, so Tesla can integrate the wiring and cooling system on the torso.
Most importantly, Optimus uses the same “navigation system” as Tesla vehicles, with Wi-Fi and LTE connectivity, which can instantly detect the movement of objects in the environment and let it go to the target without hitting the surrounding items and colleagues.
From the technical display, although Optimus has been tested in Tesla’s Fremont, California Gigafactory, it is still a long way from the actual function and mass production, which is why Musk is more eager for technical talents.
In AI Day 2022, Tesla did not show “more powerful” autopilot capabilities. Instead, this was their first detailed explanation of the operation of the autopilot software system. If you are not interested in artificial intelligence technology, you can skip this paragraph.
After Tesla releases FSD Beta in 2021, the number of car owners joining the test has increased from 2,000 to the current 160,000.
What is even more amazing is that within this year, the Tesla self-driving team has tested 75,000 software models, and released as many as 281 modules and 35 versions.
One of the most significant updates is the Occupancy Network released last month. Engineers train computers through videos instead of divided images, and successfully create automatic labels, automatic time stamps, and object trajectories.
A module that calculates the best route in as little as 10 milliseconds.
In addition, the “lane model” of FSD beta is also very different from before. Most of the “lane model” is based on the markings on the road, so the computer usually performs better on fast roads.
However, the lane of FSD beta is not only a marking, but a “reasonable driving route” automatically drawn by the computer according to the scene environment.
The role of this module is most obvious at intersections. FSD beta can now mark the starting point, middle point and end point of each lane. Even if there is no turning line at the intersection, the computer can also grasp the reasonable driving route of the vehicle. With “dynamic prediction”, the computer can complete left turns at difficult open intersections.
Importantly, such computing systems are not only used in Tesla vehicles, but also in Optimus robots, to deal with the movement in factories or homes.
It sounds simple, so why do most other manufacturers use pictures instead of videos to train self-driving systems? Because the amount of data is too large,
Ashok Elluswamy, director of Tesla’s AI team, pointed out that even though Tesla has three Dojo supercomputers and 14,000 GPUs to assist in computing, they have to process 400,000 videos per second, and the temporary storage capacity of videos is as high as 30 PB.
Powerful computing alone is not enough, and data processing, computing models and neural networks must be redesigned to do so.
The team also shared an interesting mechanism of FSD. In its neural network system, a “rating” system is set up. Through the video of the test owner, the supervisor will give a score according to the owner’s behavior, such as not yielding to pedestrians, preempting driveway, etc.
In addition, the computer also evaluates the “cost” caused by the impact, and after considering these factors, it makes the best choice.
In addition, FSDB also has the wisdom to fight against the “three treasures”. When the opposite vehicle illegally makes a left turn, although its own lane is green, the computer will evaluate the speed of the vehicle and slow down appropriately to avoid a car accident.
If there is a car owner in front of him who is playing with his mobile phone and stops on the road in a daze, the Tesla computer will automatically bypass it instead of waiting in the back.
A Tesla AI Day for Engineers
From the opening Optimus to autonomous driving, and the final appearance of the Dojo supercomputer (the space is limited, those who are interested in learning about the details of Dojo can start watching from the last half hour of the film),
Tesla AI Day 2022 seems to be deliberately flattening their organization. There are more than 20 employees who speak on the stage, including the United States, Europe, India and Chinese. It is very diverse.
Moreover, the technical details continued to be unceremoniously deepened. It can be seen that many media who were looking forward to the release of the product left early.
But this Geek speech mode is definitely more interesting for engineers who want to develop artificial intelligence or data processing than a bunch of promotional conferences piled up with marketing rhetoric.
This is exactly what Tesla wants, to show that they are strong and respect the attitude of engineers, and attract more talents to join.