Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.

Tesla Bot

This site may earn commission on affiliate links.


Active Member
Oct 2, 2013
SF Bay Area
If you want to learn how robots are programmed, then this video is a good introduction.
How robots are trained at a high level:
  1. Trained first in a virtual environment. Random movements with a reinforcement algorithm. The reinforcement algo rewards good actions and penalizes bad actions.
  2. Years of wall clock training time can be simulated in hours on the computer.
  3. After virtual environment the virtual training is applied to the real world and adjustments are made.
Apologizes: I feel there is a more appropriate existing thread for this, but I couldn't find it.
  • Helpful
Reactions: primedive
This is the better clip, from 60 minutes over a year ago:

The problem with robots is basically that you can't just deploy thousands of them "to practice". Both cost and safety issues obviously.

I'm never going to sell my GOOGL, btw.
  • Like
Reactions: DanCar
Bot moving battery cell from "conveyor belt" to tray and takes walks around the office.
Notice 43 seconds into the video where the bot was first trained by human tele operation. The dudes with the vr goggles and gloves. 49 seconds we see human tele operation with stuff and shelves.

Elon pimping the Bot: Later this year the hands will have 22 degrees of freedom.

grok: This means the hand can move and rotate in 22 different ways, allowing for a high degree of dexterity and precision in its movements. Previous had 11 DoF. The increased DoF will enable the Bot to perform more complex tasks and handle objects more delicately.
Last edited:
  • Informative
Reactions: JB47394
Narrow task robotics are better solved with an industry robot. Cheaper and faster by two orders of magnitudes. This is all smoke and mirrors. Try to disregard from the human form factor and judge the actual value produced.

its almost like the bot is still in development. This is not a sales video offering this capability. Its an update on a sample task being used to develop the way the bot is trained. Not every task is well suited to bolting a 1 ton kuka robot to a position for it to work on all day. I guess all R&D is 'smoke and mirrors' until a product is finally on sale?

edit: Oh you are the same guy spamming reddit with your insistence that the bot is useless. How depressing.
  • Like
Reactions: JHCCAZ and DanCar
I guess all R&D is 'smoke and mirrors' until a product is finally on sale?

The next decade I think is going to be really interesting in robots, the emergence of the robotics industry has been in the waiting for 10, 20 years without really emerging other than for pre-program behavior and stuff like that. And the main issue is, again, the Moravec paradox, how do we get those systems to understand how the world works and plan actions? And so we can do it for really specialized tasks. And the way Boston Dynamics [and everyone else] goes about it is basically with a lot of handcrafted dynamical models and careful planning in advance, which is very classical robotics with a lot of innovation, a little bit of perception, but it’s still not, they can’t build a domestic robot.

So until we have, again, world models, systems that can train themselves to understand how the world works, we’re not going to have significant progress in robotics. So a lot of the people working on robotic hardware at the moment are betting or banking on the fact that AI is going to make sufficient progress towards that.

There’s zero demonstration of hierarchical planning in AI where the various levels of representations that are necessary have been learned. We can do two level hierarchical planning when we designed the two levels. So for example, you have a dog-like robot, you want it to go from the living room to the kitchen. You can plan a path that avoids the obstacle, and then you can send this to a lower level planner that figures out how to move the legs to follow that trajectories. So that works, but that two level planning is designed by hand.
Last edited:
Please explain the robots playing football 1 on 1 in those terms.
They are in a bounded game world where they play each other in a simple game with a clear reward.

They train vs each other in simulation in millions of games and then the result is put into physical robots.

With “real world” robotics you cant do reward based sim training really and you can’t deploy thousands of robots to play or figure out the optimal strategy for moving forward because they break themselves and stuff and are dangerous to people if they fall or whatever. So you need a person to control and/or supervise each robot.

It’s painfully slow and costly to train them for each specific task and to gather enough data.

With FSD Tesla has millions of labeled video clips, and still can’t drive unsupervised.

If we disregard safety for a moment, Steve Wozniak is credited with coining “the coffee test” as a real test for AGI.

The test is that A robot should be able to be put in front on any random house without prior knowledge of it and go inside and make a cup of coffee. The hierarchical planning needed for that single simple task is insane. But a human does it without hardly any effort. That’s Moravec’s paradox.
Last edited:
$16K for bot , that is mostly worthless today, but lots of potential for the future > 3 years, in my opinion.
If I learned anything from buying FSD in 2019, it’s that you never buy anything based on what someone says it may be able to do later.

The hardware will be obsolete if we get to usefulness for these robots in 10-15 years.
  • Like
Reactions: DanCar