Abstract
Sensorimotor control models traditionally consist of two types of internal models: inverse models, which compute the motor commands needed to reach a desired movement goal, and forward models, which predict the resulting sensory feedback. These models are usually considered separate entities, but it is unclear whether such separation exists in the nervous system. Additionally, maintaining separate networks may be more computationally expensive. Therefore, we investigated whether these functions could be executed within a single neural circuit: an inverse-forward-recognition model (InFoRM). We implemented InFoRM using neural networks and compared their ability to reproduce cyclic reaching movements with that of control architectures based on classical, separated inverse and forward models. Desired movement trajectories were represented by recorded three-dimensional kinematics, while efferent (muscle activation) and afferent (muscle length and velocity) signals were obtained through inverse dynamics. Our findings show that InFoRM significantly outperforms control architectures across various conditions, while requiring fewer resources. The network is also able to morph to untrained movement directions, generating motor commands and predicted feedback that had not been learned. These findings demonstrate the computational advantages of integrating inverse and forward processes within a single neural network, suggesting that such unified sensorimotor models may be worthwhile to explore further.