Self-organising bio-inspired reflex circuits for robust motor coordination in artificial musculoskeletal systems.

Artificial musculoskeletal systems mimic mammalian biomechanics using antagonistic muscles and rigid skeletons. They offer benefits such as adjustable stiffness, back-drivability, and muscle failure tolerance but are difficult to model and control due to redundancies across task, joint, and muscle activation spaces, compounded by complex muscle dynamics and motion-dependent moment arms. Analytical methods require detailed system knowledge and lack scalability, while model-free approaches often rely on manual tuning and rarely exploit motor redundancy. This work introduces a model-free, biologically inspired kinematic controller based on reflex circuits that self-organise via Hebbian learning driven by Spontaneous Motor Activity (SMA). These circuits are then integrated to create a computationally inexpensive task-space controller, requiring minimal training and no analytical modelling. Simulations with six- and twelve-muscle models show that the interaction between reflex circuits, morphology, and gain modulation produces coordinated muscle synergies for human-like target reaching. Unlike previous control methods, it is easily scalable, can automatically handle unknown disturbances, and compensates for inaccessible muscles without re-training or manual intervention while maintaining high control accuracy.