ACCount39
4 days 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
4 days 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
4 days 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
3 days 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.