That's exactly the hope. Working robots have largely been out of reach for normal households and most businesses because the actuators are made in such low volumes that they are too expensive. That's why there tends to be excitement when an automaker shows interest, because they can actually possibly achieve economies of scale. Previously it was Honda, but they never went anywhere with it. We will see if Tesla actually releases an actual commercially released product.
Tasks that may require more mobility than a specialized robot can accomplish. For example a robot that takes the food to the table. Also the fact that presumably they can be trained by the owner to do tasks beyond what the company programmed them for, which allows one unit to do multiple things (instead of needing to buy multiple different specialized robots).
I don't think the initial general purpose robots may necessarily be able to accomplish that (something just based on a vague command). Rather, they may need the owner to first train them in a task then later repeat that task. However they will simply be more flexible than a specialized robot (not be stuck on rails like that fry making/burger flipping robot nor only be able to accomplish tasks it was programmed for by the company).
Do we know how many end-to-end control examples Tesla needs now? The engineering effort to scale language model training resulted in GPT-3 handling many more tasks with few-shot learning, so is there potential for a vision foundation world model to support learning from a few examples? For example a humanoid robot could be provided both good and bad examples of picking up an egg to realize what is actually an egg and that too much pressure breaks the egg.
The training for the basics would be already done (for example how to pick up an object, or how to walk from point a to point b). You just need to fine tune it with the specific task you want it to accomplish.
The legs and DOF will come for "free" (and may give some flexibility for example for bringing the items to the counter after it is done without needing another separate robot). It's the same logic of how Tesla includes hardware like heated rear seats, but have it disabled. At some point it becomes cheaper to not make so many different models even if some parts are extraneous for the given task.
The specific tasks you laid out as other pointed out is more complex than the simpler tasks that a general purpose robot would likely be initially doing.
If you can't counter a simple poke like that, perhaps its time to question your beliefs? If you think humanoid robots are imminent, please lay out the case?
Do Tesla vision foundation models need both video and current control for training? While IMU could be a decent input proxy with linear and angular velocities to correctly differentiate between waiting at an intersection or turning through it, that could be different enough from accelerator and steering wheel control.
None of my drives have even close to zero interventions. I mean, it is way better than it used to be 2.5 years ago. But it makes horrible mistakes (not necessarily true safety interventions but could increase risk) all the time.
No, he stated that the combination of the existing NNs and the C++ code was slower than an all NN based solution, so you can't deduce anything about the "unified NN" vs "stacked NN" options based on that, since we do not know the overhead of the C++ stack And NN speeds are closely tired both to depth AND breadth, so a back-end NN that reduces the breadth of the front-end control NN can actually result in a faster overall throughput than a shallower (but broader) unified NN.
I didn’t say anything about financial fraud. A lot of companies do things solely/mostly for optics. Spending a few hundred millions of this type of marketing is cheap, don’t you agree? You can book it as capex even. And get happy employees since their stock comp doesn’t crater.
It’s really quite simple: robotaxis are orders of magnitudes easier to solve than getting a humanoid to “(make and) bring me coffee and an omelette” or the car factory equivalent in a random location, even if you train it for a specific location and setup.
It doesn’t work like that. You don’t need to touch a chain to train for a chain blocking the road... You need to add lots of such chains to your training dataset and label it accordingly.
I think you read too much into what I was writing. Stock pumping is not fraud, it's talking up the prospects of a company within the limits of the law in "future looking statements", using words like "i think, if Tesla executes perfectly during the coming five years, that it could some day have a higher market cap than Apple and Aramco combined".
I can imagine a future where things work differently. But I also know how things work today and approximately where the research is at. That's being grounded in reality. The opposite is being delusional. That's why I think the "useful humanoid" is at least 10-15 years out.
Level 12 of hopium. After that we will have level 13 of hopium.
The main reason to mimic a human look in robotics is for social reasons. You don’t need a head open doors and you don’t need to be bipedal to climb stairs.
FSD Beta 10.x did allow for expanding deployment ~200x from just over a thousand private beta vehicles to ~363k. Cumulative miles similarly increased from not even a million miles to ~131M before 11.x single stack. I believe the main architectural change with 10.x was introducing memory both spatial and temporal initially with static objects including lines and road edges for better intersection predictions especially with dynamic occlusions then later applied to moving objects and occupancy network.
I find this doubtful. Current 11.4.9 has gone back to oscillatory braking/regen behavior and these hard problems seem impossible for Tesla to resolve - they just get better or worse around some local “optimal” spot which is not optimal. I’m not convinced that NNs will be the secret that solves this but I guess we’ll see.
It needs to be L3+ with a useful ODD for a significantly higher take rate.
