tested Astra’s wiring manipulation capabilities in simulation i asked it to untangle a pair of headphones with 2+1 entanglements this seemed to be its limit: with more entanglements added, it couldn’t find a workable strategy

it’s ok i can’t either

“achieved human-level capabilities on untangle earpods bench”

this rope-threading experiment inspired me to test wiring capabilities further

Any open source plan for the sim env?

here :)

impressive none the least. I'm curious is the 2+ knots a limitation of astra or the hardware?

i tried complex scenarios with more knots in sim but astra didn't manage to untangle them

Right was that because of the limitations of the arms or astra?

definitely astra, it was tractable to do with this hardware but trajectories didn’t make sene

Could you elaborate? What is Astra actually doing in this demo? Is astra generating a trajectory for both arms then executing it open loop? Is it generating end effector positions for each step and the playback is significantly sped up?

it wrote the controller that translates gripper targets into joint commands and the task code that chooses those targets for every step it wasn’t making model calls every frame, playback is 1× speed

Ahh I see. So is the model called at all during the simulation or did it just write the code the accomplish the task?

no it basically prepared it all in advance, which would be useful as rl/imitation learning data, to be tested yet how it'll perform this as a policy

Thanks for the quick responses!

you cant blame him

@SpinworksTech need cables

Interesting, could be used in data centres

@26pablo7 cracked

ONE OF US. 2+1 is solid. I need a physical rig. EarPods will hate me.

Which simulator is this?

MuJoCo

whoa, the wire simulation seems pretty good here, esp for mujoco.

ok now do the drawer with all the old chargers

that's ASI-level capabilities

Thanks for sharing this

Headphones with one extra knot is still a fair limit

The 2+1 limit is a useful result. One wrong pull can tighten another crossing and invalidate the whole plan. I’d like to see Astra reobserve the headphones after each move and plan the next pull from the updated state.

What kind of simulator is this? It simulates the cables so well!

Finding a clean limit is valuable. What failure mode showed up first with extra entanglements: perception, planning, or execution?