This study demonstrates that macaque monkeys can learn to control a brain-machine interface (BMIc) to perform reaching and grasping tasks using a robot arm. The BMIc uses multiple mathematical models to extract motor parameters such as hand position, velocity, and gripping force from the activity of frontoparietal neuronal ensembles. High BMIc accuracy required recording from large neuronal ensembles, and continuous operation led to significant improvements in both model predictions and behavioral performance. Monkeys succeeded in producing robot movements without moving their arms, indicating that the BMIc's dynamic properties were incorporated into motor and sensory cortical representations. The study addresses several fundamental issues in BMI research, including the type of brain signal to use, the number of neurons to record from, and the motor parameters that can be extracted. The results show that multiple cortical areas, including the primary motor cortex (M1), dorsal premotor cortex (PMd), supplementary motor area (SMA), posterior parietal cortex (PP), and primary somatosensory cortex (S1), contribute to BMIc performance. Single-unit activity provided slightly better predictions than multiunit activity, but multiunit activity was sufficient for high-performance BMIc operation. The study also reveals that long-term operation of the BMIc leads to functional cortical reorganization. Neuronal contributions to model predictions increased with learning, and directional tuning of neurons changed in response to BMIc operation. These changes suggest that the BMIc's dynamic properties were incorporated into cortical representations. Additionally, the study shows that the introduction of a mechanical device, such as a robot arm, in the BMIc control loop significantly impacts learning and task performance, as monkeys had to adjust to the dynamics of the artificial actuator. The findings highlight the importance of using large neuronal samples for efficient BMIc operation and suggest that the dynamic properties of the BMIc are incorporated into multiple cortical representations. The study also demonstrates that accurate real-time prediction of motor parameters can be achieved using multiunit signals, eliminating the need for complex spike-sorting algorithms. Overall, the study provides important insights into the potential of BMIc technology for restoring motor functions in paralyzed patients.This study demonstrates that macaque monkeys can learn to control a brain-machine interface (BMIc) to perform reaching and grasping tasks using a robot arm. The BMIc uses multiple mathematical models to extract motor parameters such as hand position, velocity, and gripping force from the activity of frontoparietal neuronal ensembles. High BMIc accuracy required recording from large neuronal ensembles, and continuous operation led to significant improvements in both model predictions and behavioral performance. Monkeys succeeded in producing robot movements without moving their arms, indicating that the BMIc's dynamic properties were incorporated into motor and sensory cortical representations. The study addresses several fundamental issues in BMI research, including the type of brain signal to use, the number of neurons to record from, and the motor parameters that can be extracted. The results show that multiple cortical areas, including the primary motor cortex (M1), dorsal premotor cortex (PMd), supplementary motor area (SMA), posterior parietal cortex (PP), and primary somatosensory cortex (S1), contribute to BMIc performance. Single-unit activity provided slightly better predictions than multiunit activity, but multiunit activity was sufficient for high-performance BMIc operation. The study also reveals that long-term operation of the BMIc leads to functional cortical reorganization. Neuronal contributions to model predictions increased with learning, and directional tuning of neurons changed in response to BMIc operation. These changes suggest that the BMIc's dynamic properties were incorporated into cortical representations. Additionally, the study shows that the introduction of a mechanical device, such as a robot arm, in the BMIc control loop significantly impacts learning and task performance, as monkeys had to adjust to the dynamics of the artificial actuator. The findings highlight the importance of using large neuronal samples for efficient BMIc operation and suggest that the dynamic properties of the BMIc are incorporated into multiple cortical representations. The study also demonstrates that accurate real-time prediction of motor parameters can be achieved using multiunit signals, eliminating the need for complex spike-sorting algorithms. Overall, the study provides important insights into the potential of BMIc technology for restoring motor functions in paralyzed patients.