ControlWiz 已转推
Basically you pick your poison in terms of where you want to inject human inductive bias in the loop: directly at low level (imitation) or at some abstraction level (reward shaping, initialization etc.). There are arguments for either to be more bitter lesson pilled.
One bitter lesson for RL in robotics manipulation is: it is actually often less scalable due to the amount of task-specific and environment specific tuning of each parameter to get anything working.
Yunlong Song@realyunlong
Most robotics RL paper is often just imitation learning in disguise. The "human expert" transfer task through extensive reward shaping, curricula, initialization strategies, environment design, and various tricks. You are providing demonstrations--just indirectly. A reward function is just a demonstration written in a different language.
English
