@chelseabfinn A critic that doesn't need the action spelled out is a lovely idea for robots. Curious how it could score a grasp that looked wrong and worked. Reply to @chelseabfinn: Generative value classifiers are a new parametrization of Q-functions that don't require action conditioning. We think these may combine gracefully with large, pre-trained policies and learn/generalize better from a richer output signal. Paper: https://t.co/pNphW8m0uf