Abstract

We present APT-RL(Action Pretrained Transformer based Reinforcement Learning) which achieves highspeed vision-based quadrupedal locomotion over complex terrain by encompassing two main quadruped locomotion skills, trotting and bounding. Unlike traditional pretraining-based reinforcement learning controllers, APT-RL does not rely on the acquisition of costly motion capture datasets or highly specialized policies for distinct tasks. Instead, it leverages a 2D simplified quadrupedal running dataset on flat terrain, coupled with a straightforward and cost-effective trajectory optimization technique. Our rigorous evaluation in real-world scenarios with the KAIST Hound, a 45kg full-sized quadrupedal robot, yielded experimental results including climbing over 60 centimeters high step at a speed of 4 meters per second while leveraging various gaits.

Gait change to overcome high obstacle

Indoor hurdle and high stair

Outdoor hurdle and step

Outdoor running experiment

Demonstration in front of Marc Raibert at Hubo Lab KAIST! DRCD Lab, RAI Lab