Re: Receptionist (above), which I’m fairly skeptical of
I’d love to hear your thoughts on why you’re skeptical about using Figure as a receptionist. Do you believe businesses might be hesitant to adopt it or do you see a technical limitation that prevents it from handling all the tasks a human receptionist would typically do?
I think my objections to using the Figure robot as a receptionist are mostly that -- it's $200,000 receptionist (being optimistic) that's going to be less capable than a real one, and most of the core challenges are going to be equally faced by an iPad app, just without the expensive body.
I love the video for how striking and sci-fi it looks; I think it's a nonsensical business, but that's also maybe not the point.
That totally makes sense. Unless there are sufficient capabilities (heavy-lifting hardshell luggage, operating check-in software, handling emergency procedures etc.), the use-cases shown in the demo, while cool, is a high price to pay.
"Figure always seems to keep its leg planted when doing manipulation tasks" , I am trying to think why that could be?
As I understand it, the main factor affecting this is the two system control architecture they have. Because they use two different systems to control the upper body and lower body and the two systems probably takes some assumption about the states of each other before they start. We don't have that case in Unitree as the whole body is trained as one. In case of figure the robot can not take advantage of legs for manipulation because the imitation learning system which is mainly responsible for manipulation don't have direct control over them.
Also, Since they are not controlling lower body with the imitation learning pipeline they can not record the change in position and the operator effectively needs to stay still while recording any task
or it could may as well be anything else. These are just thoughts.
Despite current humanoid seems not that good at 'intelligent decision making', I am always a believer to this human factor in LONG RUN: human tends to give more attention on what looks like them --- love or hate, you have to admit humanoid brings more imaginations.
For the near 5 years, I am skeptical whether humanoid can reach same level of manipulation skill as bimanual robot arm do (look at pi05 and DYNA). But at an exciting era, where AI grows exponentially, we may expect a 'FIGURE', whether this one or another one come out to replace it, brings generalist, whole-body control humanoid robots that can work in the wild, in your home.
I tend to agree; I think that a lot of whether a humanoid works is going to come down to integrating whole body control. Else we might as well just use wheels?
But I think Figure et al have more than 5 years of runway, now, with their massive raises; so i don't think this should block anyone. In the end it seems Waymo cleared a lot of the way by proving the value of long-term investment in real world AI
Yes WBC is key to unlock humanoid for manipulation work. I like the demo shown in Agility Robotics, our iron men should be capable of moving heavy stuff while have precise control in their hands. A 'human-like' leg may not be always necessary. Looking forward to seeing more demos!
Neat write-up!
Re: Receptionist (above), which I’m fairly skeptical of
I’d love to hear your thoughts on why you’re skeptical about using Figure as a receptionist. Do you believe businesses might be hesitant to adopt it or do you see a technical limitation that prevents it from handling all the tasks a human receptionist would typically do?
I think my objections to using the Figure robot as a receptionist are mostly that -- it's $200,000 receptionist (being optimistic) that's going to be less capable than a real one, and most of the core challenges are going to be equally faced by an iPad app, just without the expensive body.
I love the video for how striking and sci-fi it looks; I think it's a nonsensical business, but that's also maybe not the point.
That totally makes sense. Unless there are sufficient capabilities (heavy-lifting hardshell luggage, operating check-in software, handling emergency procedures etc.), the use-cases shown in the demo, while cool, is a high price to pay.
Thanks for writing this.
It's a s nice write-up!
"Figure always seems to keep its leg planted when doing manipulation tasks" , I am trying to think why that could be?
As I understand it, the main factor affecting this is the two system control architecture they have. Because they use two different systems to control the upper body and lower body and the two systems probably takes some assumption about the states of each other before they start. We don't have that case in Unitree as the whole body is trained as one. In case of figure the robot can not take advantage of legs for manipulation because the imitation learning system which is mainly responsible for manipulation don't have direct control over them.
Also, Since they are not controlling lower body with the imitation learning pipeline they can not record the change in position and the operator effectively needs to stay still while recording any task
or it could may as well be anything else. These are just thoughts.
1x also uses separate models but we don't see the same restrictions because they have a clear interface between upper and lower body
Thanks for your writing!
Despite current humanoid seems not that good at 'intelligent decision making', I am always a believer to this human factor in LONG RUN: human tends to give more attention on what looks like them --- love or hate, you have to admit humanoid brings more imaginations.
For the near 5 years, I am skeptical whether humanoid can reach same level of manipulation skill as bimanual robot arm do (look at pi05 and DYNA). But at an exciting era, where AI grows exponentially, we may expect a 'FIGURE', whether this one or another one come out to replace it, brings generalist, whole-body control humanoid robots that can work in the wild, in your home.
Let's wait until that day!
I tend to agree; I think that a lot of whether a humanoid works is going to come down to integrating whole body control. Else we might as well just use wheels?
But I think Figure et al have more than 5 years of runway, now, with their massive raises; so i don't think this should block anyone. In the end it seems Waymo cleared a lot of the way by proving the value of long-term investment in real world AI
Yes WBC is key to unlock humanoid for manipulation work. I like the demo shown in Agility Robotics, our iron men should be capable of moving heavy stuff while have precise control in their hands. A 'human-like' leg may not be always necessary. Looking forward to seeing more demos!
The recent Unitree demos out of Amazon FAR, I think, have gone the farthest in showing how cool humanlike legs can be!