Rendered at 23:05:57 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
ACCount39 23 hours ago [-]
It's impressive that something this simple can do this much. But those funky types of actuators all live and die by transfer learning now.
If a robot AI can figure out how to operate them with very little sim and teleop data, and learn to take advantage of their strengths while maintaining good performance on tasks learned from UMI datasets, teleop data or human headcam videos? Allowing the same "robot mind" to work with different actuators?
Then I would expect those to have a decent niche - sitting between the classic two finger UMI gripper and a humanoid hand. Not a drop-in replacement for a human hand, but still more dexterity per hand without sacrificing all of the ruggedness and mechanical simplicity.
If transfer learning for different actuator types doesn't work so well? I expect the field to collapse to a binary of "UMI gripper or humanoid hand", with nearly no in-between.
In general, I'm carefully optimistic? But we are yet to demonstrate with confidence that this kind of transfer for actuators with radically different kinematics would work.
measurablefunc 22 hours ago [-]
Should be doable w/ existing AIs to generate a staged sequence of operations by taking a demonstration & deriving another sequence of operations for achieving the same outcome w/ another set of actuators. I'm not a roboticist but robotic arms have specifications & those specifications are basically generalized algebraic datatypes so Astra or even Grok should be able to generate programs for manipulating most objects.
ACCount39 12 hours ago [-]
The practical gap between "specifications" and "generalized algebraic datatypes" IK and "manipulating most objects" is massive.
Motion planning is not the kind of well behaved task where you can change an actuator type and everything just works. And modern robotics specifically, the kind where demands on manipulation capability are the highest, is dealing with open ended environments - where the environment, the task and the objects involves are generally unknown in advance.
Having "strong generalists" like Astra helps bootstrap a lot of things, but that still isn't a "full solve". You don't get to go from "a series of hardcoded commands that use this actuator open a bottle in a sim" to "a set of behavioral heuristics that tell a robot AI how to use this actuator effectively for performing arbitrary operations on unseen objects" for free.
Which is why the dynamics of generalization and transfer learning between actuators are so important. If you need very little data to "bootstrap" a new actuator type, and even a few sim envs can get a robot to perform the tasks it knows from other effector embodiments with it, and start taking advantage of what the new kinematics enable? You're in a very good shape. If transfer barely works, and you need 100000 hours of real world embodied data on diverse tasks per actuator? Living hell.
measurablefunc 8 hours ago [-]
You don't need a full solve. It's a planning problem w/ formally specified operations for getting to a desired state. I wrote enough code in undergrad robotics class to know the problem isn't actually as difficult as people were saying but I didn't have enough AI at the time to do the transfer mapping between different actuators.
Animats 1 days ago [-]
It works best on problems which are strongly Cartesian. The chopsticks demo shows the limitations of this. The gripper can pick up two chopsticks, but it can't do much with them, because it can't rotate them to bring them together at the points.
But it seems to be good for lab equipment with simple geometry.
Also, where are the motors? Inside the gripper, or at the other end of cables?
octoberfranklin 1 days ago [-]
Rotational screw-on caps are not Cartesian, and this excels at dealing with them.
I'm very impressed with this.
Re: chopsticks, oh but it can! It can roll one chopstick between one pair of grippers. Do that while holding the tip against a fixed object and you'll tilt the chopstick. Once you get the tips touching, it can use them.
The motors are inside those chonky things just outside the fingertips. I'm guessing hobby servos, which are tiny cheap and strong.
varjag 1 days ago [-]
Still there are fundamental limitations. Holding a bowl full of liquid, a heavy jug with a rounded handle etc.
octoberfranklin 1 days ago [-]
What would be the problem with a bowl of liquid?
The balancing there comes from the wrist joint, not the fingers. Obviously this device is supposed to be mounted on a wrist joint that can rotate (pitch+yaw). Wrist joints are pretty well-understood these days.
Rounded handle I don't see the problem unless the hole is too small for one of the fingers.
varjag 1 days ago [-]
Holding a rounded handle with a rectangular profile slider is inherently unstable as it's extremely sensitive to the pinch point.
And the bowl has a radius that make orthogonal grip impractical.
octoberfranklin 19 hours ago [-]
But it has two pinch points. There are two separate grippers. Look at how it uses scissors.
If the radius of the bowl is that big it's too big for anything one-handed (with a reasonable size hand). So use two hands, like people do.
I feel like there could be a fifth finger at the top 90 degrees rotated compared to the others, to e.g. click a pen
octoberfranklin 1 days ago [-]
One of the challenging parts of imitating human hands is that our hands sense through the same surface which deforms as we bend our fingers. We don't have good robotic devices that can do that.
This sidesteps the problem: the surface that contacts the object is flat and never bends, so you can apply a huge variety of very detailed grid sensors to it.
The "rolling between the fingers" trick is ultimately what eliminates the need for deformation.
If a robot AI can figure out how to operate them with very little sim and teleop data, and learn to take advantage of their strengths while maintaining good performance on tasks learned from UMI datasets, teleop data or human headcam videos? Allowing the same "robot mind" to work with different actuators?
Then I would expect those to have a decent niche - sitting between the classic two finger UMI gripper and a humanoid hand. Not a drop-in replacement for a human hand, but still more dexterity per hand without sacrificing all of the ruggedness and mechanical simplicity.
If transfer learning for different actuator types doesn't work so well? I expect the field to collapse to a binary of "UMI gripper or humanoid hand", with nearly no in-between.
In general, I'm carefully optimistic? But we are yet to demonstrate with confidence that this kind of transfer for actuators with radically different kinematics would work.
Motion planning is not the kind of well behaved task where you can change an actuator type and everything just works. And modern robotics specifically, the kind where demands on manipulation capability are the highest, is dealing with open ended environments - where the environment, the task and the objects involves are generally unknown in advance.
Having "strong generalists" like Astra helps bootstrap a lot of things, but that still isn't a "full solve". You don't get to go from "a series of hardcoded commands that use this actuator open a bottle in a sim" to "a set of behavioral heuristics that tell a robot AI how to use this actuator effectively for performing arbitrary operations on unseen objects" for free.
Which is why the dynamics of generalization and transfer learning between actuators are so important. If you need very little data to "bootstrap" a new actuator type, and even a few sim envs can get a robot to perform the tasks it knows from other effector embodiments with it, and start taking advantage of what the new kinematics enable? You're in a very good shape. If transfer barely works, and you need 100000 hours of real world embodied data on diverse tasks per actuator? Living hell.
Also, where are the motors? Inside the gripper, or at the other end of cables?
I'm very impressed with this.
Re: chopsticks, oh but it can! It can roll one chopstick between one pair of grippers. Do that while holding the tip against a fixed object and you'll tilt the chopstick. Once you get the tips touching, it can use them.
The motors are inside those chonky things just outside the fingertips. I'm guessing hobby servos, which are tiny cheap and strong.
The balancing there comes from the wrist joint, not the fingers. Obviously this device is supposed to be mounted on a wrist joint that can rotate (pitch+yaw). Wrist joints are pretty well-understood these days.
Rounded handle I don't see the problem unless the hole is too small for one of the fingers.
And the bowl has a radius that make orthogonal grip impractical.
If the radius of the bowl is that big it's too big for anything one-handed (with a reasonable size hand). So use two hands, like people do.
This sidesteps the problem: the surface that contacts the object is flat and never bends, so you can apply a huge variety of very detailed grid sensors to it.
The "rolling between the fingers" trick is ultimately what eliminates the need for deformation.