Learning to Assist Humans without Inferring Rewards