AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
Modular Learning Transfers Well in Simulation
modular learning has a 90% success rate in the real world, while and when learning fails to transfer completely down 20%. classical approaches fall somewhere in between at 80% success rate. The main difference is that the end-to-end policy directly takes the RGB and depth frames as input. It's an explicit map of all the objects the robot has seen so far as input. And it performs poorly because it's faced with real world images that look very different.