ResearchBenchmarksGoal-achievement behavior naturalness on DeepMind Control Suite HumanoidFollow70.8Naturalness Preference ScoreTADPoLe62.16864.40966.6568.891Jul 2, 2024Evaluation ResultsMethodMethodLinksNaturalness Preference ScoreTADPoLePrompt=a person kneelingPrompt=a person kneeling2024.0770.8TADPoLePrompt=a person in lot...Prompt=a person in lotus position2024.0762.5TADPoLePrompt=a person doing...Prompt=a person doing splits2024.0762.5