Key Takeaways from Tesla’s Revolutionary AI Day 2022

onlyusedtesla
2022-10-01

Tesla is changing the course of history with its second revolutionary AI Day event:

Optimus Humanoid Robot

  • Fully-functioning prototype tested and demonstrated for the first time with no additional cranes or supports.
  • The robot can walk, move, wave, and operates with ten fingers.
  • Powered by the same Full Self-Driving Computer that’s in a Tesla.
  • Possible for the robot to pick up boxes, and runs on Autopilot neural networks re-trained for a robot.
  • The robot can view the world through neural networks.
  • Second prototype with Tesla-designed and built actuators, battery packs, and computer that will be similar to the production version.
  • Optimus will be able to move fingers independently, with opposable thumbs.
  • The goal is to create a humanoid robot that’s usable as quickly as possible.
  • Optimus will be built in high volumes with low cost.
  • Optimus is designed to be an extremely capable robot, made in volumes of millions.
  • Designed to cost “much less than a car, less than $20,000 would be my guess.”
  • The potential of Optimus is appreciated by very few people.
  • A lot of work needed to refine Optimus.
  • The point of AI Day is to recruit people that will bring the vision of Optimus to fruition.
  • An economy becomes quasi-infinite, creating a future of abundance with no poverty where humans can have the products and services they want.
  • Civilization will be transformed.
  • Keeping Tesla a publicly-traded company is essential to ensure that the change has a positive influence.
  • Self-driving cars will improve productivity of transport by at the most an order of a magnitude, whereas Optimus could improve economic output by two orders of a magnitude or higher — no clear limit.
  • Tesla hopes to create a transition that society wants with Optimus.
  • Original prototype was developed just six months after the concept in 2021.
  • Today’s robot is the latest generation, with a true focus on the human form.
  • 100w power consumption while sitting, 500w while walking briskly.
  • 73kg mass.
  • 200+ degrees of movement, with 27 in the hand.
  • Optimized for cost and efficiency by reducing part count and power consumption, including reducing wiring and centralizing power distribution with compute at the center.
  • 2.3kWh battery pack, 52v, with all electronics integrated within the central torso.
  • Bot Brain is in the torso, with one Full Self-Driving Computer as well as Wi-Fi, LTE, audio, and security. It can essentially do all that a human brain does.
  • 28 structural actuators, 11 degrees-of-freedoms hands.
  • Crash software from Model 3 utilized in the robot for damage control when it falls.
  • Human-inspired, non-linear joint design.
  • Optimized joint for efficient actuators.
  • The robot has 28 actuators vs 1–3 in a car.
  • Every joint has to be optimized, with too many unique actuator designs.
  • Tesla had to run commonality styling to figure out the design of actuators that could be used for multiple joints, resulting in six unique actuator designs.
  • One actuator can lift a 9-ton concert grand piano by necessitated design, as quadricep muscles could do the same.
  • The hand is a biologically-inspired design, as the world is already ergonomically-designed.
  • Six actuators within each hand, with eleven degrees of freedom.
  • The hand can carry a twenty-pound bag, use tools, and has a precision grip.
  • The neural network is the same as the network used by Autopilot, with new training.
  • Volumetric depth rendering is recognized.
  • The robot went from first steps in April 2022, to a pelvis unlock in July 2022, an arm sway in August 2022, then a toe off in September 2022 to emulate human walking.
  • Physical self-awareness is necessary for walking, as is an energy-efficient gait, balance and coordinated motion that’s all done naturally.
  • Locomotion Planning: footsteps planned first, with trajectories to connect the footsteps, then finding a mass trajectory to keep balance.
  • Reality isn’t like a simplified bot model with simplified real-world assumptions.
  • Un-modeled dynamics lead to de-stablization.
  • State estimation and motion control lead to stabilized real behavior.
  • Natural motion references from a human demonstration are used to map to the robot, with trajectory optimization to account for variation in the real-world.
  • Impressive demonstration completed in 6–8 months.
  • Ultimately, Optimus will change world economies.

Full Self-Driving

  • 2,000 customers in the Full Self-Driving Beta last year has shifted to 160,000 customers today.
  • There have been 35 releases, 281 models, 18,659 pull requests, 75,778 models trained, and a data of 4.8 million clips in between.
  • Planning software runs the Full Self-Driving system, built on training infrastructure, occupancy, lanes and objects, training data, and an AI compiler/interface.
  • A lot of objects need to be resolved within a glance, as interactions with pedestrians and vehicles need to be evaluated quickly.
  • Interaction search: parallelized tree search over maneuver trajectories.
  • Seed trajectories branch the interactions, continuously bringing in more interactions to optimize with more constraints.
  • A neural planner was built using human examples from the fleet, bringing decision times down to under 100us per action.
  • Trajectory scoring is based on collision checks, comfort analysis, intervention likelihood, and a human-like discriminator.
  • A video feed from eight cameras understands occlusions in 3D.
  • Occupancy networks: based on volumetric occupancy, multi-camera, and video context, persistency through occlusions, occupancy semantics, occupancy flow, resolution, efficient memory and compute, and runs in around ten miliseconds.
  • Dense 3D reconstruction using volumetric rendering.
  • The occupancy network is trained with a large auto-labeled dataset.
  • Video model training: requires parallel compute with three in-house supercomputers, using 10,000 GPU’s for training and 4,000 GPU’s for auto-labeling.
  • The accelerator has to be fed reliably, which is done with Tesla’s training computer Dojo.
  • Video is trained on the fleet videos directly, with no processing done before that.
  • FSD Lanes need to produce the full set lane instances, predicting their connectivity to each other.
  • FSD Lanes has its own lane detection neural network, made up of layers encoding the video streams from the vehicle enhanced by a navigation map called a lane guidance module. Tesla converts a dense tensor of lanes into a specialized language component that can encode lanes.
  • Simple approaches don’t work in many instances, thus Tesla has to develop complex systems for i.e. lane detection.
  • The networks have about a billion parameters combined, with over 1,000 neural network signals.
  • 500,000 trips per day have been cached, requiring manual and auto-labeling.
  • Automated 3D labeling by multi-trip reconstruction.
  • It would take two weeks for artists to create a rendering of an intersection, which would be too long.
  • Instead, Tesla’s system can simulate the scene in five minutes.
  • Unreal Engine finishes the project simulation, following tile creators, extractors, and loaders.
  • Data is created using weeks of work rather than months or years.
  • Tooling to identify mis-predictions, with a data engine that perfects the network.

Dojo

  • Density will drive new levels of performance for Tesla’s Dojo Accelerator.
  • The first functional training tile was demonstrated last year, with a build rate reaching one tile per day.
  • A voltage regulator module developed in-house will enable a power improvement to meet production goals.
  • All system components must be integrated into power modules.
  • Extreme levels of power — Tesla has created a supercomputer that will constantly be improved with new software and a further 10x improvement with next-generation hardware.

Q&A

  • Tesla is not just a car company: Tesla is the leader in real-world AI hardware and software, and building the most important computer in decades.
  • Open to evolving technical solutions, but the current goal is to become useful as quickly as possible with parts designed to be made at volume.
  • Optimus will be tested internally at Gigafactories.
  • “Really fun” versions of Optimus will come that can be seen as a buddy.
  • As AI improves, Optimus will gain more of a personality and the ability to understand instructions. Optimus is designed to be an android.
  • Interventions will be possible to stop the robot from damaging itself if necessary.
  • Tesla’s mission broadens to make the future awesome with Optimus.
  • Optimus will be awesome in five years, ten years will be mind-blowing.
  • Conversational capabilities will be added.
  • Localized control ROM is included as an essential safeguard.
  • Most efficient way to operate Dojo is to keep it online as an available service to train models.
  • Software 2.0 will use a lot of neural net training.
  • Elon tests alpha builds of the Full Self-Driving Beta.
  • FSD Beta will be ready for a worldwide beta by the end of the year from a technical standpoint.
  • Significant Beta improvements to be made next month.
  • Advice to younger self: enjoy the moment more.
  • The robot is an actuators and sensors problem, which Tesla is solving.
  • Optimus will start with simple tasks, with number of useful tasks growing exponentially.
  • Optimus release window within 3–5 years.
  • Tesla and SpaceX create an environment where great engineers can flourish.
  • The internet is the greatest equalizer for knowledge.
  • Elon finds right now to be the most interesting time in history.
  • Full Self-Driving won’t be perfect, but data will demonstrate it to be significantly safer than a human driver.
  • Add-ons for Optimus will be possible.
  • AI Day to be held annually with tremendous detail, and the possibility of a monthly podcast.

__________________

Interested in Listing or Upgrading YOUR Tesla? The BEST place to get TOP dollar for your used Tesla. MADE IN NYC

Tesla currently has extremely limited new inventory available for immediate purchase, and used vehicles across the automotive industry are selling for an average of $5,000 over market value.

Browse the listings now: Only Used Tesla

Buy or list your premium electric vehicle: OnlyEV

Or feel free to contact us directly at contact@onlyusedtesla.com today!

Disclaimer: Investing carries risk. This is not financial advice. The above content should not be regarded as an offer, recommendation, or solicitation on acquiring or disposing of any financial products, any associated discussions, comments, or posts by author or other users should not be considered as such either. It is solely for general information purpose only, which does not consider your own investment objectives, financial situations or needs. TTM assumes no responsibility or warranty for the accuracy and completeness of the information, investors should do their own research and may seek professional advice before investing.

Comments

  • Gregho
    2022-10-02
    Gregho
  • Brrrrrrrrrrr
    2022-10-02
    Brrrrrrrrrrr
    Half the world seems interested in destruction only. Offer something good, yet they dont want to see it come to fruition.
  • JaiVenky
    2022-10-02
    JaiVenky

    Great effort 👌😊

  • TKPK
    2022-10-02
    TKPK
    thanks for sharing 👍
  • Snakewood
    2022-10-02
    Snakewood
    😱😱😱👏👏👏👍👍
  • mark01bravz
    2022-10-02
    mark01bravz
    Optimus will end poverty!
Leave a comment
59
3