A Continual Learning Method for Generalized Grasping Manipulation in a Musculoskeletal Robot

Musculoskeletal robotic systems offer structural advantages while presenting significant control challenges. Current research on their manipulation capabilities, particularly multi-object grasping scenarios, remains insufficient. Furthermore, as robots operate in dynamic environments with evolving task requirements, developing their ability to grasp novel objects while maintaining existing manipulation skills becomes crucial. To address these challenges, we propose a novel continual learning method for generalized grasping manipulation in a musculoskeletal upper limb robot. First, we propose an end-to-end learning method for multi-object grasping that leverages object-specific latent features and multi-level states of a musculoskeletal robot. Additionally, a novel continual learning method for grasping tasks is proposed with a bio-inspired object-preference-based experience replay selector, which optimizes learning efficiency while mitigating catastrophic forgetting. Our approach successfully masters the grasping manipulation task of 10 objects in the first phase and continually learns 23 additional objects in the second phase, outperforming existing methods in both multi-object grasping and continual learning. Furthermore, our analyses of muscle synergy and methodological robustness demonstrate that the proposed approach generates biologically plausible muscle synergies and exhibits strong robustness against object observation bias and neural excitation noise. Our experiments, conducted both in simulation and on a hardware system, demonstrate the practical transferability of our method to an articulated dexterous hand grasping task. Note to Practitioners—Musculoskeletal robots, which incorporate human-like joints, muscles, and actuation mechanisms, exhibit exceptional robustness, dexterity, and compliance, offering promising applications and significant control challenges. While current research predominantly addresses lower limb locomotion and upper limb movements, studies on multi-object grasping and continual learning remain limited. This article presents a novel continual learning method for generalized grasping manipulation in a musculoskeletal upper limb, enabling end-to-end multi-object grasping without the dependency on predefined or generated grasping postures. The proposed methodology encompasses two key components: 1) the extraction and processing of object point cloud features as partial observations for the controller and 2) the implementation of a bio-inspired object-preference-based experience replay selector that balances new object learning and existing skill retention. Experimental evaluations reveal that the proposed approach exhibits effective muscle synergy patterns and demonstrates robust performance under diverse noise conditions and magnitudes, indicating significant potential for real-world deployment and cross-domain adaptation. Nevertheless, future work should address current limitations, including the extension from fixed-point to arbitrary position grasping, and the advancement from single-step grasping to complex multi-stage manipulation and assembly tasks.