<ul><li>Quadrupedal locomotion via Reinforcement Learning (RL) is commonly addressed using the teacher-student paradigm.</li><li>Teacher-Aligned Representations via Contrastive Learning (TAR) is proposed to bridge the representation misalignment between the privileged teacher and the student.</li><li>TAR achieves robust generalization to Out-of-Distribution (OOD) scenarios, outperforming fully privileged methods.</li><li>TAR allows adaptive locomotion and continual fine-tuning in real-world scenarios without requiring privileged states.</li></ul>

TAR: Teacher-Aligned Representations via Contrastive Learning for Quadrupedal Locomotion

Discover more