Evaluating Data Augmentation Method for Command-Conditioned Hexapod Locomotion in Uneven Terrain

We investigate reinforcement learning (RL) for command-conditioned locomotion of hexapod robots in uneven terrain using simulated environments. Due to their high degrees of freedom, hexapods are prone to local optima during learning, resulting in suboptimal policies. To address this, we apply a data augmentation method called Virtual Command Allocation (VCA), which replaces commands in collected experiences with randomly sampled ones from the command space. Simulation results across multiple terrain conditions show that VCA improves the stability and generalization of learned policies, enabling robust and adaptive locomotion. These findings demonstrate the effectiveness of VCA for learning command-conditioned control even in challenging and variable environments.