Optimus already sorting objects autonomously | TMC Podcast #51

hooty · Sep 24, 2023

Umm, Optimus isn't coming.
The failure of 'FSD' is proof

Bitdepth · Sep 24, 2023

Similar research from Toyota.

THEbuz · Sep 25, 2023

hooty said:
Umm, Optimus isn't coming.
The failure of 'FSD' is proof

How have you concluded that FSD is a failure? What evidence of this claim can you point to? To anyone paying attention, it is clear that FSD continues to advance with fewer and fewer interventions. I’m not sure how you can see that as anything but progress. Sure, Elon’s timeline assertions have been way off, but that does not mean that FSD is a failure. I need to see more evidence to support your claims.

Dewg · Sep 25, 2023

THEbuz said:
How have you concluded that FSD is a failure? What evidence of this claim can you point to? To anyone paying attention, it is clear that FSD continues to advance with fewer and fewer interventions. I’m not sure how you can see that as anything but progress. Sure, Elon’s timeline assertions have been way off, but that does not mean that FSD is a failure. I need to see more evidence to support your claims.

It's okay Buz - there are a few people here to whom we just say "there, there" and offer them a hot beverage.

Jonty3 · Sep 25, 2023

hooty said:
Umm, Optimus isn't coming.
The failure of 'FSD' is proof

I mean i've seen plenty of videos of FSD driving around some complex cities so, it's not really a failure

THEbuz · Sep 25, 2023

Jonty3 said:
I mean i've seen plenty of videos of FSD driving around some complex cities so, it's not really a failure

Reminds me of one well-intended Teamster with whom I had previously worked at UPS. He drove tractor trailers—often triples—for UPS. When we were discussing the impending and eventual release of fully autonomous transport, he said, “No computer can ever do what I do,” to which I thought to myself, “Famous last words.” Sure, these guys pull triples and double-45ft trailers, often in snowy and icy conditions, but those challenges are not insurmountable; they are solvable in the real world, in terms of FSD. I’m not anti-union (I was 3x a member of various unions), but it is important for the uninitiated to learn where the term Luddite comes from.

spacecoin · Sep 26, 2023

A CV sorting video in a controlled environment would have been impressive 10 years ago. These narrowband demos are meaningless for a humanoid robot, which is hyped as a replacement for humans.

This soccer in sim is more impressive to me (from last year):

At the end of the day this is a research area and meaningful products are likely 10+ years out.

FSD is a failure in terms of what was promised and its current rate of improvement. It doesn’t look like it will be able to operate without a human in the loop for the coming five years in a meaningful ODD.

THEbuz · Sep 26, 2023

spacecoin said:
A CV sorting video in a controlled environment would have been impressive 10 years ago. These narrowband demos are meaningless for a humanoid robot, which is hyped as a replacement for humans.

This soccer in sim is more impressive to me (from last year):

At the end of the day this is a research area and meaningful products are likely 10+ years out.

FSD is a failure in terms of what was promised and its current rate of improvement. It doesn’t look like it will be able to operate without a human in the loop for the coming five years in a meaningful ODD.

I dunno. I’m unconvinced. I wouldn’t be surprised if we are seeing L5 by sometime in 2025.

jeewee3000 · Sep 26, 2023

spacecoin said:
FSD is a failure in terms of what was promised and its current rate of improvement. It doesn’t look like it will be able to operate without a human in the loop for the coming five years in a meaningful ODD.

Five years is a very long time in AI development. By that time Tesla will be on HW5 at least (they indicated 3 years per HW suite, in reality it was around 4 years IIRC), compute will have exploded and E2E will have been finetuned a lot.

Also, breakthroughs from competitors are analyzed by Tesla AI/X.ai and incorporated if necessary.

Compare today to five years ago. The data collection rate from Tesla has exponentially increased. And data is everything.

JulienW · Sep 26, 2023

spacecoin said:
A CV sorting video in a controlled environment would have been impressive 10 years ago. These narrowband demos are meaningless for a humanoid robot, which is hyped as a replacement for humans.

This soccer in sim is more impressive to me (from last year):

At the end of the day this is a research area and meaningful products are likely 10+ years out.

FSD is a failure in terms of what was promised and its current rate of improvement. It doesn’t look like it will be able to operate without a human in the loop for the coming five years in a meaningful ODD.

Yes, but I believe the difference is it was not given any rules, instructions or training on how to do it but was just "told" to do it. So it figured out the "rules" on its own. That is a big difference if so and shows it could have real time adaptability to all kinds of unknown situations.

uscbucsfan · Sep 26, 2023

THEbuz said:
How have you concluded that FSD is a failure? What evidence of this claim can you point to? To anyone paying attention, it is clear that FSD continues to advance with fewer and fewer interventions. I’m not sure how you can see that as anything but progress. Sure, Elon’s timeline assertions have been way off, but that does not mean that FSD is a failure. I need to see more evidence to support your claims.

I'm not saying it's a failure, yet, but your claim of fewer and fewer interventions isn't supported by the community sourced data. Some versions are better than others, but 11.4.7 is a bit worse than 10.69.2, which was worse than 10.11.2. They are re-writing a ton and Elon claims much of the V12 AI is in 11.4.7, so regressions are expected, but the data doesn't show gradual improvement with disengagements...it shows peaks and valleys with a mean of about the same for the last 14 or so months.

We are no where near the point of the March of 9s and we will for sure see some additional regressions with the additional re-writing. It's difficult to say if it will ever be level 5 at this rate.

spacecoin · Sep 26, 2023

THEbuz said:
I dunno. I’m unconvinced. I wouldn’t be surprised if we are seeing L5 by sometime in 2025.

I would be surprised if we see the use of computer vision unsupervised in radiology in five years. Driving with solely computer vision in a wide ODD? I seriously doubt it in the coming five years. Level 5? Not in 20 years.

JB47394 · Sep 26, 2023

JulienW said:
Yes, but I believe the difference is it was not given any rules, instructions or training on how to do it but was just "told" to do it. So it figured out the "rules" on its own. That is a big difference if so and shows it could have real time adaptability to all kinds of unknown situations.

If "told" means "trained by example" and "rules" means "heuristics", then I would agree with you.

I figure they put a guy in that training suit and repeated the sorting task a few dozen times. That was enough information for the neural networks to glean how to pick up, move and drop blocks as well as understand where the blocks should go. The resiliency of the robot would depend on the diversity of the training. For example, if they never moved the trays or swapped sides, then the robot may well have learned nothing more than "blue goes to the right" and "green goes to the left". Swap the trays or move them and you may not get the result you were after.

I would be most interested in what it takes to train a task with very few variants versus a long task with many. They're not going to train an Optimus robot to drive a car using this technique.

I'll include a Tweet by Julian Ibanz, a Tesla engineer

Super excited to share our team's progress over the last few months. We can now accomplish long horizon tasks such as sorting blocks fully autonomously in a completely task-agnostic way.

Just collect more data and we can train a new complex task without changing any code!

That suggests to me that if they had a task to push a button and flip a switch at a different orientation, they'd have somebody back in that suit pushing a button and flipping a switch a few dozen times. That Optimus robot would know how to do that, but not pick and place colored blocks. It's not yet a general solution, but a task-specific framework.

zoomer0056 · Sep 26, 2023

Dewg said:
"there, there" and offer them a hot beverage.

Ok Sheldon,

.

spacecoin · Sep 26, 2023

jeewee3000 said:
Five years is a very long time in AI development. By that time Tesla will be on HW5 at least (they indicated 3 years per HW suite, in reality it was around 4 years IIRC), compute will have exploded and E2E will have been finetuned a lot.

Also, breakthroughs from competitors are analyzed by Tesla AI/X.ai and incorporated if necessary.

Compare today to five years ago. The data collection rate from Tesla has exponentially increased. And data is everything.

Waymo was autonomous five years ago and are making progress - but it's not quick. It's super hard work getting even a 2x in 18 months by perhaps the smartest team in the business.

I think people underestimate the reliability needed for autonomy. Tesla has barely doubled their (super low) disengagement rate in three years - from 5-10 city miles per DE to 10-20. Even if you 10x that over the next three years FSD will still be at least 200x from autonomy.

Data is definitely not everything. Autonomous driving would have been solved years ago if that was true.

It turns out that five years isn't that long. The "Attention is all you need" (transformers) paper and things like AlphaGo was over SIX years ago and its not like it's a completely different world now. Radiology is still a supervised CV application.

Not much has happened in terms of fundamental research breakthroughs since that. Computer vision (alone) is likely not safe enough for unsupervised safety critcial applications until we get a few more of those kinds of breakthroughs (in particular around explainability and robustness).

I'm more bullish on other ML application areas such as in biotech, medicine and personal assistants.

hooty · Sep 26, 2023

THEbuz said:
How have you concluded that FSD is a failure? What evidence of this claim can you point to? To anyone paying attention, it is clear that FSD continues to advance with fewer and fewer interventions. I’m not sure how you can see that as anything but progress. Sure, Elon’s timeline assertions have been way off, but that does not mean that FSD is a failure. I need to see more evidence to support your claims.

Just need to go over Musk’s quote history on FSD, going back to 2014, to know that the term ‘progress’ is overstated

hooty · Sep 26, 2023

Jonty3 said:
I mean i've seen plenty of videos of FSD driving around some complex cities so, it's not really a failure

I tend to use objective evidence as opposed to anecdotes - in no state does Tesla have a licence to test

hooty · Sep 26, 2023

jeewee3000 said:
Five years is a very long time in AI development. By that time Tesla will be on HW5 at least (they indicated 3 years per HW suite, in reality it was around 4 years IIRC), compute will have exploded and E2E will have been finetuned a lot.

Also, breakthroughs from competitors are analyzed by Tesla AI/X.ai and incorporated if necessary.

Compare today to five years ago. The data collection rate from Tesla has exponentially increased. And data is everything.

Why don’t they have a robotaxi after 9 years now?

BBTX · Sep 27, 2023

Comparing the relatively low-key skills that Optimus has shown off so far to other "research" isn't really a valid comparison. Everyone else is really just doing it as a research project. If they're lucky, they'll finally get one working example of a better-than-Optimus robot engineered, and it will be not be reproducible at scale for reasonable cost. Even if they get there, they then have to re-engineer everything to scale, and it will probably turn out that a lot of the decisions made in research don't pan out at scale.

What Telsa's doing with Optimus is going in from the start with a goal of being able to cheaply manufacture this at scale. It's a whole different ballgame. Every time they succeed at some small goal like sorting objects, they're not competing with some research project that did it in a lab once a decade ago. They're in an empty playing field, where they're arguably the only company that will be able to mass-produce what they're showing off at a reasonable price.

Optimus already sorting objects autonomously | TMC Podcast #51

Member

Member

Member

Active Member

New Member

Member

Active Member

Member

Active Member

Well-Known Member

Active Member

Active Member

Active Member

Active Member

Active Member

Member

Member

Member

Member

Similar threads