Researchers release TAGA to extend AME2, using active gaze and sparse attention to cut humanoid locomotion training costs to 20%
The method improves robot agility across challenging terrains
Positive users highlight how AME2 follow-ups enable emergent gaze and sparse attention to cut legged robot training costs to 20 percent while allowing expansion to new use cases, whereas negative users express disappointment.
Most Activity
i've been doing this myself. you can think of this as foveated rendering that VR headsets use. most of the time gets taken up rendering, but what you can do is actually just give the policy the ability to "direct the eye" and it learns really well in RL
Such a cool line of work
I've had policies learn interesting things like scanning behaviours without me explicitly having them learn it. Ray tracing is very batchable too. This was deployed to real
i've been doing this myself. you can think of this as foveated rendering that VR headsets use. most of the time gets taken up rendering, but what you can do is actually just give the policy the ability to "direct the eye" and it learns really well in RL
:(
@yacineMTB It's always better to have someone else do the work.
@chris_j_paxton Emergent gaze from RL alone. That's the key bit.
@yacineMTB on the flip side. You can take their work and expand it to your use case
@yacineMTB did you mean to publish or leverage for your co products?
if pub, yea sorry (but you can still go further i guess)
but if prod then all the more PoC to double-down
consider who invented the mouse, icons/windows, DOS, see who ended up selling it en masse neither Xerox nor IBM