[Mr. Roboto]

With AI Day happening, let's have this thread be the spot for all things Optimus Bot and AI announcements!

 

[Administrator]

 

[JPforReal]

The vibes ahead of the live stream are way better than the Cyber Rodeo

 

[JPforReal]

Musk said it would start right at 6:15pm and he wasn't lying. The crowd didn't seem ready when he walked out!

 

[JPforReal]

Promising updates on Autopilot and Dojo, plus a "long Q&A"

And with that, the robot is out - walking and moving all on its own, apparently for the first time ever without any supports. Talk about confidence. They must have been nervous!

 

[JPforReal]

"The robot can do a lot more than that, we just didn't want it to fall on its face." jokes Musk.
Then they rolled some video of the robot watering plants in an office and doing work in one of Tesla's plant. The second shot is apparently showing what the robot sees.
One of the engineers pointed out it uses the same computer as the Tesla vehicles do for Autopilot.

 

[JPforReal]

Those demos were of a development robot, not the version they expect to produce at scale. Next they've brought out a rough version of the Optimus robot they actually expect to put into production. Musk is saying it should cost less than $20,000.
It's clearly a lot less polished than the development unit we saw before. You can see the three guys here who helped move it out on the stage while Optimus waved.
A few minutes later, something happened off camera and there was some awkward laughter. Then they rolled a somewhat slumped Optimus off stage.

 

[JPforReal]

Now Musk is talking about the potential for productivity increases, while making sure its done responsibily. He's reiterating an earlier point about Tesla being a good company to do this because it's publicly traded and if people don't like what they're doing, they can get involved by buying shares and change the company's direction, or even fire him.

 

[JPforReal]

They're talking now about how quickly they've moved since last year's joke with a person in a robot suit.
It seems they've done a lot by focusing on a minimum scope of mobility - not trying to absolutely recreate what the human body can do.

 

[JPforReal]

Musk promised on Twitter earlier that this would be a very technical briefing and he wasn't lying. They're talking now about design with cost and manufacturing efficiencies in mind, with battery capacity and runtime estimates.

 

[JPforReal]

New engineer on stage now talking about how they've used their vehicle crash modelling to design the robot to be able to fall down and take abuse without significant damage. There's an adorable animation of the robot doing a faceplant.
(EDIT - the "sensitive content" is a screengrab of an animated Model 3 crash test, incase you were worried)

Sensitive content, not recommended for those under 18

Show Content

 

[JPforReal]

There was just mention of "Angular Contact Ball Bearings" and I'm reminded that Musk said this briefing would be technical with the intention of recruiting engineers - I think the point of all this is to prove their serious about what they're doing here and if you're on the fence, this might spur you to want to get involved.

 

[JPforReal]

The actuator they've designed can lift a grand piano. THAT I understand!

 

[JPforReal]

They've mimiced a human hand because factories are desinged for human bodies. So it was easier to match workflows that were designed for human ergonomics than to expect everything in the world to be redesigned for the robot

 

[JPforReal]

The vision system is lifted straight from Autopilot. The bot needs to be able to navigate on its own, including finding recharge stations when needed.

 

[JPforReal]

Next up: walking. It's really hard.
Physical self-awareness is something we take for granted as humans, but needs to be recreated entirely for robots. The bot has to plan and calculate each footstep

 

[JPforReal]

This was all done, they say, in six or eight months.
Now we're on to Full Self Driving for Tesla's cars. After reviewing the number of models they rolled out in a year, and developing the neural network, they're now showing how they develop the decision making. Like with the bot, it's something humans do well but takes a lot of work to model for computers.
The neural network allows them to make the various scenario modelling much faster, which enables running through the dozens of options for the car in any moment, in real time. It's hard to see, but in the last slide they show how they went from 1-5ms per calculation to 100us.

 

[JPforReal]

They're talking now about the computing power needed to process all the video models they've created to train the FSD system. It's absolutely staggering the scope of the tools and systems they've needed to design in the background just so they could get to work designing the systems (like FSD) we know about.