This is the same training technique Stanford ALOHA is using<p><a href="https://mobile-aloha.github.io/" rel="nofollow">https://mobile-aloha.github.io/</a><p><a href="https://www.trossenrobotics.com/aloha.aspx" rel="nofollow">https://www.trossenrobotics.com/aloha.aspx</a><p><a href="https://venturebeat.com/automation/stanfords-mobile-aloha-robot-learns-from-humans-to-cook-clean-do-laundry/" rel="nofollow">https://venturebeat.com/automation/stanfords-mobile-aloha-ro...</a>
Pretty obvious deception, IMHO.<p>If the point is to demonstrate the training (which is copying a human) then why hide the human off screen? Because it creates the impression that it's autonomous.<p>Putting "but it's not autonomous" in a subsequent tweet is just plausible deniability.<p>His fans will argue this is just a simple misunderstanding. C'mon guys. You used to be smarter than this.
The text of his tweet also "gives away the magic trick":<p>> Optimus cannot yet do this autonomously, but certainly will be able to do this fully autonomously and in an arbitrary environment (won’t require a fixed table with box that has only one shirt).
If this can be used for training then this magic trick can be enough. Training a language model uses in principle also just one magic trick - trying to predict a next sentence in a large set of ordered sentences.
> Optimus doesn’t appear have capabilities beyond anything we could do in 1964<p>Seems a bit disingenuous. Did the Lincoln robot have the dexterity in its fingers to do fine-motor tasks like folding laundry?<p>This has the same energy as saying the Falcon 9 isn't a massive innovation because the Saturn V existed in the 60s. Propulsive landing was done back in the 60s by the Apollo LEM, so obviously the propulsively landing re-usable first stage was "already done before".