Abstract
Abstract
Humanoid musculoskeletal robots driven by McKibben muscles offer inherent compliance and safe physical interaction, making them ideal for human-centered environments. However, developing the capacity for learning and generalizing reaching movements to new spatial targets remains a complex mission due to the sophisticated coordination required between perception, decision-making, and control mechanisms. Inspired by the human motor system, we propose a brain-inspired framework for motion learning and generalization on a musculoskeletal robotic arm, integrating two key neural principles: motor primitive encoding and cerebellum-inspired predictive correction. We first learn reusable motor primitives from sparse demonstrations using a multilayer perceptron (MLP) optimized via iterative imitation learning. To generalize to novel targets, we propose a Cerebellum-Inspired Analytic Motor Primitive Composition (CAMPC) method that integrates a differentiable forward model and muscle-synergy-aware regularization to compute optimal blending weights. Experiments show that our CAMPC method reduces mean positioning error to 2.45 cm in simulation and generalizes to hardware using human demonstrations without any reinforcement learning, thereby avoiding hazardous online policy exploration on fragile pneumatic actuators. On the physical robot, CAMPC attains the lowest mean error of 3.03 cm, outperforming other baselines (paired t-tests, p<0.05). This demonstrates that our brain-inspired method enables humanoid musculoskeletal robots to generalize to untrained static reaching targets with high accuracy and data efficiency within a constrained desktop workspace.