Continuous Skill Learning for the ANYmal Robot

Abstract

This work presents a novel framework for continuous skill learning tailored for the ANYmal quadruped robot. The proposed approach leverages the principles of skill decomposition, recombination, and refinement to enable the robot to efficiently acquire, adapt, and generalize new motor skills by building on past experiences. To achieve this, various sequence models, including Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and state-space models, are explored for encoding skills within a continuous latent space. Additionally, a Partial State-Conditioned Flow Adapter is introduced to structure the latent space, enabling more efficient exploration and skill selection during policy learning. The core contribution of this work is the development of a two-tier policy architecture, which integrates a skill latent policy that uses previously learned skills with a refinement policy designed to fine-tune and adapt these skills to new tasks and environments. Empirical validation is conducted across a diverse set of terrains, both seen and unseen during training, demonstrating the robot’s ability to refine its skills dynamically and efficiently. This research has the potential to advance the field of continuous learning in robotics, providing new insights into lifelong learning, skill representation, and the potential applications of structured latent spaces for enhancing robotic adaptability and performance in complex environments.