Learning Robot Locomotion from Diverse Datasets

Quadrupedal robots have garnered much attention in recent years. Meanwhile, the Generative Pre-trained Transformer (GPT) models have achieved remarkable success in natural language processing. The abundance of online datasets offers a promising avenue to extend this framework to robotic motion by representing motion sequences as tokenized data. In this work, we retarget a diverse dataset of dog, horse, and motion data from other robot platforms onto Unitree Go2/A1, and employ a GPT-style network to generate motion sequences of arbitrary length conditioned on given gait and duration. With a low-level policy, Unitree Go2 can demonstrate diverse and natural gaits in simulation.