GROVE: A Generalized Reward for Learning Open-Vocabulary Physical Skill

GROVE Figure 1 . Open-vocabulary physical skills learned with our generalized reward framework GROVE. The white humanoid mannequins demonstrate diverse skills from abstract instructions (e.g., "conduct the orchestra," "position body in a shape of C") using a pre-trained controller. The system also generalizes to standard RL benchmarks, including the Ant, Humanoid, and quadrupedal ANYmal, all without task-specific reward engineering.