May the Force Be with You: Biomimetic Grasp Force Decoding for Brain Controlled Bionic Hands

Intracortical brain-computer interfaces (iBCIs) have achieved remarkable progress in restoring arm and hand movement by inferring motor intent from neural signals in primary motor cortex (M1) and realizing the intended movements in a bionic limb. However, manual interactions with objects require not only restoration of movement but also the precise application of forces on objects, which implies a different mode of limb control that has been largely overlooked. One of the major obstacles in incorporating force control is the lack of understanding of how manual forces are encoded in M1 during object interactions. To fill this gap, we recorded the neural activity in M1 as monkeys grasped sensorized objects with varying levels of force. We found that static decoders could not reliably extract force information from M1 activity, suggesting a dynamic relationship between force and neural activity. Consistent with this hypothesis, a recurrent neural network could exploit these dynamics to accurately decode time-varying forces. Next, we applied the insights gleaned from our experiments with able-bodied macaques to build decoders of manual force in a human participant with tetraplegia. First, we found that the patterns of responses in human M1 during imagined force application were similar to those in monkey M1 during physical force application. We then applied recurrent neural networks to decode force from M1 activity and showed that these allow the participants to accurately exert forces with a (brain-controlled) virtual hand.