3 ms·

Humanoid robots might be hard but they've been successfully demonstrated. It might also be argued that machine learning can make complex motions easier than if

by elurg 5y ago

Humanoid robots might be hard but they've been successfully demonstrated. It might also be argued that machine learning can make complex motions easier than if everything is hand-coded. The humanoid form factor might be an entirely unnecessary distraction, a robot with more legs or non-human articulations would be just as useful.

542458 5y ago
Boston dynamics has found machine learning to be minimally useful here, and has had better results hand-coding most things. That’s not to say that is necessarily correct for all scenarios, but BD has lots of extremely smart and motivated people and I would take their results seriously - so at very least it doesn’t seem to be as easy as throwing a big neural net at it, and any machine learning approach will require hard and careful work. Re: successfully demonstrated, again Boston Dynamics might be the best example. Their bipeds are nowhere near the functionality levels Elon is describing - fine motor skills, vision, even walking (it’s my understanding that Boston Dynamics has to individually tune their bipeds for every course) are years and years away from hitting Elon’s goals at current rate of progress. Furthermore, the BD bipeds have a battery life of about ten minutes. You need a revolution in either energy storage density or kinematic efficiency or more likely both to get a useful assistant bot out of it.
- JohnJamesRambo 5y ago
  I hadn’t even thought of the physics constraints on the energy needed. This is a truly ridiculous goalpost he has set unless he has developed the arc reactor to put in its chest.
ClumsyPilot 5y ago
"Humanoid robots might be hard but they've been successfully demonstrated." None of these robots can hang our laundry to dry, de-bone a chicken or cook a meal. They cannot perform any of the thousands of tasks we perform every day that require dexterity. They have just about mastered locomotion, and that's it.

by elurg 5y ago

542458 5y ago

Boston dynamics has found machine learning to be minimally useful here, and has had better results hand-coding most things. That’s not to say that is necessarily correct for all scenarios, but BD has lots of extremely smart and motivated people and I would take their results seriously - so at very least it doesn’t seem to be as easy as throwing a big neural net at it, and any machine learning approach will require hard and careful work. Re: successfully demonstrated, again Boston Dynamics might be the best example. Their bipeds are nowhere near the functionality levels Elon is describing - fine motor skills, vision, even walking (it’s my understanding that Boston Dynamics has to individually tune their bipeds for every course) are years and years away from hitting Elon’s goals at current rate of progress. Furthermore, the BD bipeds have a battery life of about ten minutes. You need a revolution in either energy storage density or kinematic efficiency or more likely both to get a useful assistant bot out of it.

JohnJamesRambo 5y ago
I hadn’t even thought of the physics constraints on the energy needed. This is a truly ridiculous goalpost he has set unless he has developed the arc reactor to put in its chest.

ClumsyPilot 5y ago

"Humanoid robots might be hard but they've been successfully demonstrated." None of these robots can hang our laundry to dry, de-bone a chicken or cook a meal. They cannot perform any of the thousands of tasks we perform every day that require dexterity. They have just about mastered locomotion, and that's it.