Seamless multi-skill learning: learning and transitioning non-similar skills in quadruped robots with limited data.

Abstract

In multi-skill imitation learning for robots, expert datasets with complete motion features are crucial for enabling robots to learn and transition between different skills. However, such datasets are often difficult to obtain. As an alternative, datasets constructed using only joint positions are more accessible, but they are incomplete and lack details, making it challenging for existing methods to effectively learn and model skill transitions. To address these challenges, this study introduces the Seamless Multi-Skill Learning (SMSL) framework. Integrated within the Adversarial Motion Priors framework and incorporating self-trajectory augmentation techniques, SMSL effectively utilizes high-quality historical experiences to guide agents in learning skills and generating smooth, natural transitions between them, addressing the learning difficulties caused by incomplete expert datasets. Additionally, the research incorporates an adaptive command sampling mechanism to balance the training opportunities for skills of various difficulties and prevent catastrophic forgetting. Our experiments highlight potential issues with baseline methods when imitating incomplete expert datasets and demonstrate the superior performance of the SMSL framework. Sim-to-real experiments on real Solo8 robots further validate the effectiveness of SMSL. Overall, this study confirms the SMSL framework's capability in real robotic applications and underscores its potential for autonomous skill learning and generation from minimal data.