TCLL: Temporal-Aware Contrastive Latent Learning for Heterogeneous Legged Robots Locomotion

Recent research on legged robot locomotion control had made significant progress, particularly in reinforcement learning approaches that demonstrated remarkable capabilities in complex environments. However, current robot reinforcement learning for locomotion control still relied on privileged information during training. Although such information accelerated convergence and raised the performance ceiling, it was unavailable at deployment, thereby widening the sim-to-real gap. This study presented a novel locomotion control method for legged robots, temporal-aware contrastive latent learning (TCLL), which operated using only limited proprioceptive data. TCLL enabled legged robots to navigate challenging terrains without relying on privileged data. To validate the effectiveness of TCLL, we conducted comprehensive simulations and physical robot experiments. The results showed that our control method outperformed existing algorithms, thereby advancing the development and practical application of legged robots in complex environments.