Changes to task relevance in redundant motor environments distinguish structure learning from adaptation

Motor actions often exceed the dimensionality of task feedback, which creates control redundancy that requires learners to identify which movement dimensions influence task outcomes. The computations that shape learning in redundant tasks, especially when task-relevant spaces change, remain unclear. We studied this by combining established visuomotor perturbation approaches with a redundant motor interface reaching task. Two monkeys learned a 3D hand position to 2D cursor position mapping that allowed us to define and manipulate the task-relevant space. After animals learned an initial task-relevant space, we introduced two types of perturbations: Relevance Preserving (RP), which rotated hand-cursor relationships within the learned task relevant space, and Relevance Altering (RA), which altered the task relevant space. We hypothesized that RP perturbations would engage error-based adaptation (observed in non-redundant tasks), whereas RA perturbations would engage additional learning mechanisms associated with acquiring new skills. RP perturbations were learned rapidly during training (online) and generalized to the initial learned mapping, consistent with error-based adaptation. In contrast, adapting to the first RA perturbation took longer, engaged learning outside of training (offline), and did not generalize to the initial learned mapping. Animals also adapted faster across repeated exposures to new RA mappings and changed how they generated movements to learn each new mapping, consistent with computations associated with structure learning. These findings suggest that changing task relevance shifts motor learning from adaptation toward structure learning, which has implications for models of motor learning and sensorimotor therapies like brain-computer interfaces.

Most real-world motor tasks provide many ways to achieve a goal. This redundancy requires learning the task's structure, including which actions are relevant, but many studies do not incorporate redundancy. We added redundancy to a visuomotor perturbation paradigm to selectively perturb task-relevant spaces. Perturbed visual-motor mappings that preserved task-relevant spaces produced rapid adaptation, while mappings that altered task-relevant spaces yielded signatures of structure learning, including history-dependent meta-learning. Our study reveals a continuum between forms of learning typically studied separately and demonstrates that cognitive inferences about task structure shape exploration in redundant motor tasks. Our work highlights the need to consider structure learning in therapies like brain-computer interfaces, and advances our ability to study the computations involved in real-world motor skills.