#488kache@YACINEMTBI've been deploying some new general control policies on drones, and I got so pissed off with calibrating one of my sensors I just randomized its orientation in the training sim
And the AI can still use it to fly..
#488kache@YACINEMTBIt learned this odd little calibration behavior where it oscillates in a circle, on roll and pitch axes while very slowly yawing left. Really fascinating
kache@yacineMTBI've been deploying some new general control policies on drones, and I got so pissed off with calibrating one of my sensors I just randomized its orientation in the training sim And the AI can still use it to fly..
#488kache@YACINEMTBI think the thing that really excites me, when it comes to robotics, is where state "estimation" is directly improved by the behaviour itself. And then, the behaviour can be improved by the improved estimation..
kache@yacineMTBIt learned this odd little calibration behavior where it oscillates in a circle, on roll and pitch axes while very slowly yawing left. Really fascinating
#488kache@YACINEMTBthis drone is being flown by a generalist neural network, with compute on the drone. The compute can tolerate any orientation of the sensors. You'll notice it do a little wobble. It's a learned strategy to calibrate everything in its hidden state. This will work on any drone
kache@yacineMTBI'm building the Canadian DJI with a debit card
#488kache@YACINEMTBThe compute is very cheap. Cheaper than you can possibly imagine. A dishwasher has enough compute for this.
See the broken sensors on one of the axes? It doesn't matter. The neural network can handle it in the fusion. Just teach it to handle broken sensors
kache@yacineMTBthis drone is being flown by a generalist neural network, with compute on the drone. The compute can tolerate any orientation of the sensors. You'll notice it do a little wobble. It's a learned strategy to calibrate everything in its hidden state. This will work on any drone
#488kache@YACINEMTBThis is a test for new infrastructure, which drives down inference cost greatly, and therefore, control hz
I have policies that can do backflips on command. Give me a week
kache@yacineMTBThe compute is very cheap. Cheaper than you can possibly imagine. A dishwasher has enough compute for this. See the broken sensors on one of the axes? It doesn't matter. The neural network can handle it in the fusion. Just teach it to handle broken sensors
#488kache@YACINEMTB
kache@yacineMTBThis is a test for new infrastructure, which drives down inference cost greatly, and therefore, control hz I have policies that can do backflips on command. Give me a week