[somenewaccount1]

people are saying this is "fake" because the overall motions were programmed ahead of time.

However, picking stuff up perfectly, placing it, and jumping, just using sensors - which is what this is really illustrating - is not that fake to me.

I could imagine this robot being used by airlines to move baggage. A confined set of pre-programmed movements with the only variable being the luggage on a cart, and a human supervisor. It reduced the back-breaking task of moving luggage to a just a normal "stand there and press buttons" type of job. Easier to higher for, maintain that crew member, and less injuries to baggage handlers.

[ericmcer]

Giving it a narrow set of pre-programmed movements defeats the entire purpose of building a human shaped robot. The goal is for it to be general purpose. If they wanted to perform a single discrete task they could design a way more efficient machine for doing that task.

[Cyph0n]

The complexity involved - across hardware, sensors, and software - is mind boggling, even for this “dumb”, pre-programmed robot. So to me, it makes complete sense to build towards the ultimate goal in a gradual manner.

[colordrops]

You need to build basic skill primitives for the robot until it's no longer a mechanical problem, and just a software problem, which can be iterated on much more quickly and take advantage of mounds of ML/AI advances. Seems like they are getting close.

[93po]

Exactly. This demo is about the foundations of motor control and dexterity of human-like movements within a human form factor.

[alpaca128]

How else would you iterate towards a more general purpose robot? You have to start with narrow tasks. Humans are general purpose but we don't teach them to drive cars before they can walk.

[dahfizz]

But you could upload a new program of movements on the same hardware.

Requiring bespoke hardware for every possible task is like saying we shouldn't have CPUs - if you're gunna write a program, might as well put it on an FPGA.

[smat]

Boston dynamics also published a making of video. There it is explained why the luggage scenario is still far away: The robot has perfect knowledge of the object it is transporting (size and weight distribution). In order to be able to transport any object, the control algorithms need to be able to estimate themselves where to grip an object and what the objects weight distribution is.

Still, the results are very impressive.

[ikekkdcjkfke]

It just needs to be fluent in moving in physical space. Tesla made a "language model" by converting traffic intersections to a language and running gpt-3 stuff

[kaba0]

It’s “programmed” as in you tell a kid to go there and grab that thing and come back.

Actually executing those motions have a ridiculous amount of complexity, hell, standing in one place and not falling over is far from trivial in case of such a robot.

[senectus1]

its just as fake or real as Tesla's self drive mode.

which is to say that its a long way from being perfectly reliable, but its also quite usable in many scenarios.

[marcusverus]

I think Atlas could see mainstream adoption without the high reliability that's expected from Autopilot. Atlas still needs to reliably not kill people, but this is seemingly much easier to do when you're not hurtling down the road at 70 miles per hour. Beyond that, it could be much less reliable than humans and still add tremendous value.

[cactusplant7374]

A broken clock is right twice a day. An unsupervised FSD enabled car crashes once per month? Probably more.

[sekai]

Wait till you hear how that compares to human driving.

[cactusplant7374]

Human drivers crash once per day? Sorry, there is no comparison. The only reason you don't see FSD crashes is because FSD testers so far have been very attentive (they are better drivers than the general public because they are enthusiasts that bought a luxury car).

But if FSD was unsupervised it would be a killing machine.

[kaba0]

If you think of yourself as a good driver, how often did you crash?

[jes5199]

how often do human drivers crash?

[lll-o-lll]

I try to keep my crashes below once a month ;)

[throwaway9915]

I am sure if you counted for example Skoda Octavia crashes it'd be way more than once a day.

Let's talk about "adaptive cruise control" too - I am not aware of any company other than Tesla having any sort of working crash protection built into that feature. My old Ford Mondeo just makes a loud noise and brakes a little, but you're going to crash anyways, it's not going to move the steering wheel an inch.

Volvo cars with "lane assist" will consistently turn you into the ditch - because it greatly overcompensates when you get near the line, which can be deadly above 100 km/h. And it's pretty hard to turn off that feature while driving and it turns on automatically when you power on the car.

Why do people keep hating on the provably most safe car (Tesla Model S) instead of mentioning all the other cars with terrible unsafe assistants? Roads would be much safer place if everyone had a Tesla.

[cactusplant7374]

I've never crashed my car driving in downtown Seattle. But FSD cannot handle downtown Seattle or downtown Chicago from the videos seen on YouTube. The crashes would be at least one per day. It's even unacceptable for a system such as this to crash the car once per year when it's a $40,000+ purchase.

[hunta2097]

There is another video showing the making-of this video.

They are very open about how much effort it took to stage, it isn't a completely dynamic interaction.

https://www.youtube.com/watch?v=XPVC4IyRTG8