Additional from Dave and John:
Occupancy network
Roadster rides (possibly SpaceX package)
Also will likely be a V11 FSD timeline and details regarding one NN.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Tesla AI day
- Thread starter Buckminster
- Start date
Occupancy network
Roadster rides (possibly SpaceX package)
Also will likely be a V11 FSD timeline and details regarding one NN.
Imo this can be deconstructed. AI and robotics engineers knows what problems need to be solved. Tesla has solved some of them or have ideas that will lead to solving them in the near future.
So what needs to be solved?
1. Perception, ie how the the world around the robot looks like
2. High level reference, ie what is the task that is to be done, for example "Locate the box, pick it up, put it down in a different place"
3. Finding the optimal high level control sequence to perform the task, ie how should the joints move at each time step to accomplish this goal
4. Setting the low level controller ie the voltage etc to get the joints to the desired angles(has been solved)
1. Is just FSD stack but for a different 3D environment
2. Can be split into to parts.
2.a Human having a nice interface to tell the robot what's to be done
2.b Robot interpreting and generating a sequence of tasks that accomplishes this goal
3. The cost function from previous AI day extended, like an MPC but for a much more nonlinear system using AI to evaluate the cost function and search heuristic.
4. Just normal stuff like inverse kinematics, PID, MPC, optimal control etc. Everyone already knows how to do this, Spacex has mastered this and Tesla knows this very well.
Xepa777
Banned
Random: do we k ow what chinas AI team is doing? I remember when tesla opened up its Shanghai r&d center that Elon said the Chinese would develop their own AI approach. Was that scrapped?
Was it 'own approach' or 'own data center'? I thought the issue was camera data leaving the country.
Tesla unveils its humanoid robot for 'less than $20,000'
Tesla has unveiled its Optimus humanoid robot at its AI Day 2022 today and Elon Musk believes Tesla can bring it to market for “less than $20,000”. As expected, the event started with Tesla unveiling a working prototype of its humanoid robot – a project first announced at Tesla’s AI day in 2021...
electrek.co
explains the panel gaps.
I look forward to seeing that long video of Tesla bot doing things they will show after AI day.
Think HW4 is a bit hyped up. HW3 is fine, HW4 can handle slightly larger neural networks so performance on some tasks might go up a bit, but software improvements will likely have a larger effect so that's more important but not as easy to imagine.
Dojo 2 is interesting, will be interesting to hear what their plans are beyond 100x in 2years. And we will see how the industry shapes up with AMD entering it in force now and China are probably not super happy about being left behind, so we will see what they do.
I expect a lot of AI day 2023 will be about scale. Scaling the infrastructure, scaling autolabel, scaling their neural networks, scaling bot production, scaling bot learning etc.
Think HW4 is a bit hyped up. HW3 is fine, HW4 can handle slightly larger neural networks so performance on some tasks might go up a bit, but software improvements will likely have a larger effect so that's more important but not as easy to imagine.
Dojo 2 is interesting, will be interesting to hear what their plans are beyond 100x in 2years. And we will see how the industry shapes up with AMD entering it in force now and China are probably not super happy about being left behind, so we will see what they do.
I expect a lot of AI day 2023 will be about scale. Scaling the infrastructure, scaling autolabel, scaling their neural networks, scaling bot production, scaling bot learning etc.
Robots did that 8 years ago:Maybe Tesla Bot will be able to drive the Tesla cars and we can forget about FSD.
I wonder when a Tesla bot will be able to drive a car better than a human. Don't think it's that much more difficult than making FSD super safe. Sure its vision will not be as overkill as HW4 sensor suite, but my vision is not that great either and I can do a decent job of it most of the time.
Last edited:
Similar threads
